Method and device for prolonging snapshot sight distance of long-distance wide-angle electronic police

Abstract

The invention provides a method and device for prolonging the snapshot sight distance of a long-distance wide-angle electronic police, and relates to the technical field of data processing, and the method comprises the steps: calculating the space volume distribution of a vehicle in a monitoring region according to a pyramid geometric model, and generating a three-dimensional space state parameter set containing the space attitude, motion vector and volume characteristics of the vehicle; and according to the three-dimensional space state parameter set, calculating vertex coordinates of a vehicle motion path, determining a final imaging opportunity of the vehicle at the key monitoring point, and generating a final snapshot control instruction sequence including a vehicle head, a vehicle body middle part and a vehicle tail. The method can effectively improve the snapshot efficiency of the key law violation of the large truck, achieves a good off-site law enforcement application effect, greatly reduces the occurrence rate and the decedent rate of major traffic accidents, and creates a safe, orderly and smooth road traffic environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a method and device for extending the shooting distance of long-distance wide-angle electronic police. BACKGROUND

[0002] With the continuous improvement of the intelligent level of urban traffic management, the electronic police system, as an important technical means for road traffic safety law enforcement and order management, has been widely applied. The traditional electronic police system usually relies on fixed cameras deployed at intersections to realize automatic detection and shooting of traffic violations such as running a red light, crossing a line, and not driving in a guided lane through image analysis of video streams.

[0003] However, the prior art has some technical challenges in actual application, especially in complex urban intersections, long-distance monitoring, and large-angle coverage scenarios: Some methods are based on two-dimensional image analysis of single frames or limited consecutive frames. This method is difficult to accurately restore the real position, posture, and motion trajectory of the vehicle in three-dimensional space. Due to factors such as perspective distortion, vehicle occlusion, and light changes, it is difficult to accurately determine whether the vehicle has crossed the stop line or occupied a specific lane based on two-dimensional image features such as bounding boxes and pixel displacement, especially when large vehicles partially occlude small vehicles or vehicles are in the edge distortion area of the image, the accuracy is difficult to guarantee.

[0004] For example, the vehicle triggers the shooting as soon as it touches the line. This method cannot adapt to different speeds, vehicle types, and motion states of vehicles. For high-speed vehicles, it may trigger too early, and the vehicle in the shooting image has not yet fully entered the illegal state; for low-speed or stopped vehicles, it may trigger too late, missing key evidence, and it is difficult to ensure that the key moment images of the vehicle head, middle, and tail with the most legal evidence value can be captured in all cases. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a method and device for extending the shooting distance of long-distance wide-angle electronic police, which can improve the accuracy of automatic detection of traffic violations.

[0006] To solve the above technical problems, the technical solutions of the present application are as follows: In a first aspect, a method for extending the shooting distance of long-distance wide-angle electronic police, the method comprising: obtaining a sequence of consecutive images of a vehicle at a red light intersection, the sequence of consecutive images including a vehicle head image, a vehicle middle image, and a vehicle tail image; preprocessing the sequence of consecutive images to obtain preprocessed image data; calculate a key contour chord length parameter of the vehicle in the imaging plane from the preprocessed image data; construct a perspective projection model according to the key contour chord length parameter in combination with camera calibration parameters; determine three-dimensional reconstructed vehicle spatial position data according to the perspective projection model; construct a pyramid geometry model of a vehicle occupancy space based on the three-dimensional reconstructed vehicle spatial position data; calculate a spatial volume distribution of the vehicle in the monitoring area according to the pyramid geometry model, and generate a three-dimensional spatial state parameter set containing vehicle spatial posture, motion vector and volume characteristics; calculate vertex coordinates of a vehicle motion path according to the three-dimensional spatial state parameter set, determine a final imaging timing of the vehicle at a key monitoring point, and generate a final snapshot control instruction sequence containing a vehicle head, a vehicle middle and a vehicle tail; control a synchronous camera array deployed at the intersection to trigger and capture key perspective images of the vehicle head, the vehicle middle and the vehicle tail at a predetermined time point according to the final snapshot control instruction sequence.

[0007] In a second aspect, a device for extending a snapshot viewing distance of a long-distance wide-angle electronic police is provided, and the device comprises: An acquisition module is configured to acquire a continuous image sequence of a vehicle at a red light intersection, the continuous image sequence comprising a vehicle head image, a vehicle middle image and a vehicle tail image; and perform preprocessing on the continuous image sequence to obtain preprocessed image data. A calculation module is configured to calculate a key contour chord length parameter of the vehicle in the imaging plane from the preprocessed image data; construct a perspective projection model according to the key contour chord length parameter in combination with camera calibration parameters; and determine three-dimensional reconstructed vehicle spatial position data according to the perspective projection model. A construction module is configured to construct a pyramid geometry model of a vehicle occupancy space based on the three-dimensional reconstructed vehicle spatial position data; calculate a spatial volume distribution of the vehicle in the monitoring area according to the pyramid geometry model; and generate a three-dimensional spatial state parameter set containing vehicle spatial posture, motion vector and volume characteristics. A determination module is configured to calculate vertex coordinates of a vehicle motion path according to the three-dimensional spatial state parameter set; determine a final imaging timing of the vehicle at a key monitoring point; and generate a final snapshot control instruction sequence containing a vehicle head, a vehicle middle and a vehicle tail. A judgment module is configured to control a synchronous camera array deployed at the intersection to trigger and capture key perspective images of the vehicle head, the vehicle middle and the vehicle tail at a predetermined time point according to the final snapshot control instruction sequence.

[0008] The above scheme of the present application has at least the following beneficial effects: By constructing a pyramid geometric model and accurately calculating the spatial volume distribution, the technical problem that the traditional electronic police is difficult to completely capture the overall appearance of large vehicles (such as buses and trucks) in long-distance monitoring is effectively solved, and the images of the vehicle head, the middle part of the vehicle body and the vehicle tail can be clearly obtained, thereby providing complete evidence chain for illegal judgment.

[0009] Based on the key contour chord length parameter and the perspective projection model, the three-dimensional reconstruction technology breaks through the limitation of traditional two-dimensional image processing, realizes accurate spatial position, attitude and size measurement of the vehicle in the monitoring area, significantly improves the positioning accuracy, effectively reduces the law enforcement errors caused by position misjudgment; by analyzing the three-dimensional space state parameter set and calculating the vehicle motion path vertex coordinates, the system can intelligently predict the best snapshot opportunity, overcome the problem of improper snapshot opportunity (too early or too late) in the prior art, greatly improve the effective snapshot success rate, and reduce the invalid image processing burden. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 is a flowchart of a method for extending the snapshot viewing distance of a long-distance wide-angle electronic police provided by an embodiment of the present application.

[0011] Figure 2 is a schematic diagram of a device for extending the snapshot viewing distance of a long-distance wide-angle electronic police provided by an embodiment of the present application.

[0012] Figure 3 is a schematic diagram of a support structure provided by an embodiment of the present application.

[0013] Figure 4 is a comparison schematic diagram of the snapshot viewing distance of a test intersection provided by an embodiment of the present application.

[0014] Figure 5 is a picture of a vehicle passing through an electronic police before the extension support is installed and the lens is replaced provided by an embodiment of the present application.

[0015] Figure 6 is a picture of a vehicle passing through an electronic police after the extension support is installed and the 8mm lens is replaced provided by an embodiment of the present application.

[0016] Figure 7 is a picture of a vehicle passing through an electronic police after the extension support is installed and the 8mm lens is replaced provided by an embodiment of the present application.

[0017] Figure 8 is a picture of a truck actually violating the law captured by an electronic police provided by an embodiment of the present application. DETAILED DESCRIPTION

[0018] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in many forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0019] As shown in Figure 1 An embodiment of the present application proposes a method for extending the capture view distance of long-distance wide-angle electronic police, which comprises the following steps: Step 1, acquiring a continuous image sequence of a vehicle at a red light intersection, the continuous image sequence including a vehicle head image, a vehicle middle image and a vehicle tail image; Step 2, pre-processing the continuous image sequence to obtain pre-processed image data; Step 3, calculating a key contour chord length parameter of the vehicle in the imaging plane based on the pre-processed image data; constructing a perspective projection model according to the key contour chord length parameter and camera calibration parameters; determining three-dimensional reconstructed vehicle spatial position data according to the perspective projection model; Step 4, constructing a pyramid geometry model of the vehicle occupied space based on the three-dimensional reconstructed vehicle spatial position data; calculating the spatial volume distribution of the vehicle in the monitoring area according to the pyramid geometry model, and generating a three-dimensional space state parameter set including vehicle spatial posture, motion vector and volume characteristics; Step 5, calculating the vertex coordinates of the vehicle motion path according to the three-dimensional space state parameter set, determining the final imaging timing of the vehicle at the key monitoring point, and generating a final capture control instruction sequence including the vehicle head, the vehicle middle and the vehicle tail; Step 6, controlling the synchronous camera array deployed at the intersection to trigger and capture key view angle images of the vehicle head, the vehicle middle and the vehicle tail at the predetermined time point according to the final capture control instruction sequence.

[0020] In this embodiment, the continuous image sequence of the vehicle head, the vehicle middle and the vehicle tail is acquired simultaneously to record the whole process of the vehicle passing through the intersection; the pre-processed standardized image data can accelerate the subsequent feature extraction and three-dimensional reconstruction calculation process; the accurate mapping from two-dimensional image to three-dimensional space is realized, the calculation accuracy of the vehicle position coordinates in three-dimensional space is greatly improved through the combination of the key contour chord length parameter and the camera calibration parameter; the pyramid geometry model can accurately reflect the real volume distribution of the vehicle in three-dimensional space, accurately calculate the relative position relationship between the vehicle and the stop line, reduce the misjudgment rate of illegal judgment, and generate a three-dimensional space state parameter set (including spatial posture, motion vector and volume characteristics) to provide a comprehensive decision basis for the dynamic capture strategy; the best capture timing is dynamically predicted based on the three-dimensional space state parameter, solving the technical problems of too early capture or too late capture of the traditional system.

[0021] In a preferred embodiment of the present application, step 1, obtaining a continuous image sequence of the vehicle at the red light intersection, the continuous image sequence includes vehicle head image, vehicle middle image and vehicle tail image, including: A wide-angle high-definition camera array is deployed at the red light intersection to synchronously collect video streams of vehicles passing through the intersection at a preset high frame rate. The camera array includes at least one long-focus camera and one wide-angle camera, wherein the long-focus camera is used to capture vehicle detail features, and the wide-angle camera is used to cover the entire intersection monitoring area, ensuring that the entire process of the vehicle from entering the monitoring range to leaving is recorded; a vehicle detection algorithm is used to identify moving targets in the video stream in real time and determine the type of target vehicle. When the vehicle enters the preset trigger area, a multi-frame continuous acquisition mode is automatically started, and the first frame image of the vehicle head, the second frame image of the vehicle middle and the third frame image of the vehicle tail are captured in time sequence to form a continuous image sequence containing the complete spatial position change of the vehicle. At the same time, the time stamp, camera number and spatial coordinate information corresponding to each frame of image are recorded, and the image sequence is bound with the unique identifier of the vehicle and stored in the buffer area for subsequent processing. This process ensures the continuity of the image sequence in time and the completeness of the coverage in space.

[0022] In a preferred embodiment of the present application, step 2, pre-processing the continuous image sequence to obtain pre-processed image data, including: The quality of the collected continuous image sequence is evaluated, and low-quality image frames caused by motion blur, overexposure or underexposure, occlusion, etc. are removed, and valid images meeting the clarity and contrast threshold are retained. Then, the valid images are geometrically corrected, and then the corrected images are color space converted and normalized, the images are uniformly converted to a standard color space, and the brightness and contrast are adjusted to a preset range to eliminate the influence of different lighting conditions on the consistency of the images. Further, background difference method and image segmentation technology are used to extract the vehicle foreground area from each frame of image, remove the background interference information irrelevant to the vehicle, and generate a binary mask image with the vehicle as the core target. Finally, the mask image and the original image are registered and cropped to obtain pre-processed image data containing only the vehicle target and standardized in size, and the pre-processed image data is reorganized into a standardized image sequence in time sequence.

[0023] In a preferred embodiment of the present application, step 3, calculating the key contour chord length parameters of the vehicle in the imaging plane from the pre-processed image data, including: Step 300, performing edge detection operator on the pre-processed image data to extract the closed boundary curve of the vehicle contour, specifically including: According to the normalized image data obtained after preprocessing, an edge detection operator based on gradient amplitude is used to calculate the brightness gradient of the image in the horizontal and vertical directions pixel by pixel. Then, the gradient amplitude image is processed by a non-maximum suppression algorithm to retain the local maximum points in the gradient direction to refine the edge response. Next, a double-threshold hysteresis connection algorithm is applied to determine the strong edge points with gradient amplitude higher than the high threshold value, and the weak edge points with gradient amplitude between the high and low threshold values, and only those weak edge points connected to the strong edge points are retained as valid edges. After obtaining the preliminary edge pixel set, the system performs edge tracking and connection operations to connect the edge pixels with adjacent spatial positions and consistent gradient directions into continuous edge line segments by scanning the entire image space. Finally, endpoint analysis and closure judgment are performed on all edge line segments to ensure that the vehicle target outer contour forms a complete, uninterrupted closed boundary curve through interpolation or extension operations.

[0024] Step 301, fitting the closed boundary curve as an elliptical geometric model and determining the major axis direction and major and minor axis parameters of the ellipse, specifically including: The vehicle target closed boundary curve extracted in step 300 is uniformly sampled to accurately obtain the image plane two-dimensional coordinates of 200 contour points. The geometric center of all sampling points is calculated as the initial center point of the ellipse: specifically, the x-coordinate values of all 200 sampling points are summed and the sum is divided by the total number of sampling points 200 to obtain the x-coordinate of the ellipse center point; similarly, the y-coordinate values of all 200 sampling points are summed and the sum is divided by 200 to obtain the y-coordinate of the ellipse center point. Based on the geometric center point calculated in this way, the coordinate offset of each sampling point relative to the center is calculated, i.e. the x and y coordinates of each sampling point are subtracted from the x and y coordinates of the center point, respectively, to generate a two-dimensional offset vector for each point. Finally, all 200 offset vectors are arranged in the sampling order to construct a 200x2 offset vector matrix, which completely and quantitatively represents the spatial distribution pattern and dispersion degree of the contour point set around the center point.

[0025] Based on the above constructed offset vector matrix, its 2x2 covariance matrix is calculated, which contains the distribution variance of the contour points in the x and y directions of the image coordinate system and the covariance information between them. Then, the eigenvalue decomposition is performed on the covariance matrix to obtain two eigenvalues and their respective eigenvectors. The two eigenvalues calculated are arranged in descending order of numerical value, and the eigenvalue with the largest numerical value corresponds to the first principal component of the contour point distribution. The direction of the corresponding eigenvector represents the direction in which the contour extends most strongly in the two-dimensional plane. The eigenvector is normalized and the angle between it and the horizontal axis (x-axis positive direction) of the image coordinate system is calculated. This angle is determined as the major axis direction angle of the ellipse. Correspondingly, the eigenvalue with the second largest numerical value corresponds to the second principal component, and the direction of the normalized eigenvector corresponding to it is determined as the minor axis direction of the ellipse which is strictly orthogonal to the major axis direction.

[0026] After the major axis direction and the minor axis direction are determined, the ellipse axis length is calculated by coordinate projection. First, for the major axis direction: a local coordinate system is established with the determined ellipse center point as the coordinate origin and the major axis direction vector as the reference axis. The coordinates of all 200 original contour sampling points are converted to this local coordinate system, and the projection coordinate value of each point in the major axis direction is calculated. The projection coordinates of all points are traversed to find the maximum point and the minimum point. The Euclidean distances between these two extreme points and the ellipse center point are calculated respectively, and the larger one of the two distance values is taken and multiplied by 2 to determine the length of the major axis of the ellipse.

[0027] Then, for the minor axis direction: another local coordinate system is established with the ellipse center point as the origin and the minor axis direction vector as the reference axis, and the projection coordinates of all sampling points in this minor axis direction are calculated. Similarly, the maximum point and the minimum point of the projection coordinates are found, their distances to the center point are calculated, and the larger distance value is taken and multiplied by 2 to determine the length of the minor axis of the ellipse. Thus, the system obtains the initial complete parameter set of the ellipse model, including the center point coordinates, the major axis direction angle, the major axis length and the minor axis length.

[0028] Step 302, selecting multiple equally spaced sampling points on the vehicle contour along the major axis direction, and calculating the Euclidean distance between adjacent sampling points as the local chord length, specifically including: The length of the major axis of the ellipse calculated in step 301 is used as a reference scale to determine the spatial range of sampling. Specifically, the line segment of the ellipse from the theoretical left end point to the theoretical right end point along the major axis direction is evenly divided into 14 equal intervals, thereby defining 15 sampling section positions. The length of each interval is equal to the total length of the major axis of the ellipse divided by 14, thereby ensuring that the subsequent sampling points are evenly distributed along the vehicle longitudinal direction (major axis direction) and fully cover the entire contour range from the front to the rear of the vehicle.

[0029] Next, a one-dimensional main axis coordinate system is established: the center point of the ellipse determined in the above step 301 is taken as the coordinate origin, and the positive direction of the main axis is taken as the positive direction of the coordinate axis. In this coordinate system, the position coordinates of the 15 sampling sections are accurately calculated according to the equal-interval principle determined in the first stage. Among them, the coordinate of the first sampling point corresponds to the left end point (the minimum coordinate value) in the direction of the main axis, the coordinate of the last sampling point corresponds to the right end point (the maximum coordinate value), and the coordinates of the middle 13 sampling points are calculated in strict accordance with the equal-interval relationship to form an ordered position coordinate sequence.

[0030] For each sampling point position coordinate calculated in the second stage, a straight line passing through the point and strictly perpendicular to the direction of the main axis is constructed as the sampling section line of the position. Then, each sampling section line is calculated for geometric intersection with the original vehicle closed boundary curve extracted in step 300; through an efficient numerical algorithm, the two intersection point coordinates of each section line and the contour curve are accurately solved, and the two intersection points are located on the two sides of the vehicle contour at the section, and the specific spatial orientation (such as up and down or left and right) is determined by the actual posture of the vehicle in the image and the imaging angle.

[0031] After obtaining each pair of intersection point coordinates, the two-dimensional Euclidean distance between the two intersection points is calculated, and the distance value is defined as the local contour chord length corresponding to the current sampling section position, which directly and accurately quantifies the imaging width of the vehicle at the specific longitudinal position. At the same time, the three core information of each sampling point is structured, associated and recorded: the position coordinate of the point in the main axis coordinate system, the corresponding pair of contour boundary intersection point coordinates, and the calculated local chord length value. Finally, the associated information of all 15 sampling points is organized and summarized according to their position order in the direction of the main axis, forming a complete and structured local chord length sampling data set.

[0032] Step 303, the local chord length is associated with the ellipse geometric parameters to generate a global chord length feature vector representing the vehicle contour shape, the global chord length feature vector includes the width distribution, length proportion and contour curvature variation characteristics of the vehicle on the imaging plane, specifically including: With the length of the minor axis of the ellipse determined in step 301 as the reference scale, the 15 local chord length values calculated in step 302 are divided by the length of the minor axis respectively, obtaining a set of normalized relative width values. This operation eliminates the absolute scale difference caused by the different actual sizes of vehicles and imaging distances. At the same time, the absolute position coordinates of each sampling point in the major axis direction recorded in step 302 are normalized with the length of the major axis of the ellipse as the reference, and the proportional coefficient of each sampling point relative to the length of the major axis is calculated, thereby obtaining a set of relative position sequences independent of absolute position. At this point, the original local chord length data is converted into a scale-normalized width sequence and its corresponding normalized position sequence.

[0033] Secondly, based on the normalized width sequence, the width distribution feature extraction is performed. The core statistical quantities of the width sequence are calculated, including the maximum value, the minimum value, the arithmetic mean value and the standard deviation in the sequence; the maximum value corresponds to the widest part of the vehicle profile on the imaging plane, the minimum value corresponds to the narrowest part, the average value reflects the average width level of the profile, and the standard deviation quantifies the dispersion degree of the width along the major axis direction. These statistical quantities together constitute the first feature subset describing the global width distribution feature of the vehicle.

[0034] Then, the profile curvature variation feature analysis is performed, and the normalized width sequence is differentiated to obtain a first-order difference sequence and a second-order difference sequence along the major axis direction. The first-order difference sequence represents the rate of change of the vehicle profile width, and the extreme points thereof correspond to the regions where the profile width changes rapidly; the second-order difference sequence represents the acceleration of the width change rate, and can be used to identify the bending or inflection point regions of the profile. Further, the zero-crossing point and extreme point distribution of the first-order difference sequence are analyzed, and the variance of the second-order difference sequence is calculated. These parameters together constitute the second feature subset describing the smoothness and curvature variation features of the profile.

[0035] Subsequently, the profile regionalization shape feature calculation is implemented. According to the normalized position sequence, the entire vehicle profile is divided into three feature sub-regions in the major axis direction: a front region corresponding to the vehicle head, a middle region corresponding to the vehicle body, and a rear region corresponding to the vehicle tail. The average value of the normalized width values in each sub-region and the average value of the absolute values of the first-order difference in the sub-region are calculated respectively. The average width and its change rate of the vehicle head region can reflect the inclination features of the front bumper and the hood; the average width and its stability of the vehicle body region can reflect the regularity of the vehicle cabin structure; the average width and its change trend of the vehicle tail region can reflect the modeling features of the trunk and the rear bumper. The shape parameters of the three regions together constitute the third feature subset.

[0036] Finally, multi-feature fusion and vector generation are performed. The three feature subsets extracted in the above steps, i.e., the width distribution statistical features, the contour curvature variation features, and the regionalized morphological features, are sequentially spliced and combined. Specifically, according to a preset dimension order, the width statistics (maximum value, minimum value, mean value, and standard deviation), the curvature variation parameters (the number and position of the first-order difference extreme points, and the second-order difference variance), and the regional morphological parameters (the width mean values and the change rate mean values of the vehicle head, body, and tail) are sequentially arranged. Finally, the system generates a structured and multi-dimensional global chord length feature vector. The vector comprehensively and compactly encodes the contour morphological essential information of the vehicle on the imaging plane, including the width distribution law, the contour curvature variation characteristics, and the morphological details of the key parts, thereby providing key and robust two-dimensional image feature descriptions for subsequent steps of high-precision three-dimensional space reconstruction combined with camera calibration parameters.

[0037] In a preferred embodiment of the present application, a perspective projection model is constructed according to the key contour chord length parameters combined with the camera calibration parameters, including: In step 304, a pre-calibrated camera intrinsic matrix is obtained, which includes focal length parameters, principal point coordinates, and lens distortion coefficients, specifically including: A calibration parameter file pre-stored for the camera deployed at the current red light intersection is read. These core parameters are accurately measured and stored during the camera installation and debugging stage by using the standard Zhang Zhengyou chessboard calibration method, and constitute the basis for subsequent geometric calculations.

[0038] The read calibration file is parsed, and a camera intrinsic matrix is extracted and constructed, which specifically includes the following parameter groups: Focal length parameters: including horizontal focal length fx and vertical focal length fy, both in units of pixels. These two parameters define the distance from the optical center of the camera lens to the imaging plane, and are key to determining the imaging scale and the two-dimensional to three-dimensional space mapping scale.

[0039] Principal point coordinates: including principal point offsets cx and cy. The coordinate point defines the intersection of the camera optical axis and the imaging plane, i.e., the position of the origin of the image coordinate system on the actual pixel plane, which is usually close to the image center but has a small pixel-level offset due to physical errors in lens and sensor assembly.

[0040] Lens distortion coefficients: including parameters for correcting image geometric distortion, mainly including radial distortion coefficients (k1, k2, k3) and tangential distortion coefficients (p1, p2); the radial distortion coefficients describe the barrel or pillow distortion of image points deviating from their ideal positions along the radial direction due to the shape of the lens; the tangential distortion coefficients describe the tangential offset distortion caused by the non-parallelism of the lens and the imaging plane.

[0041] Step 305, perform tensor product operation on the global chord length feature vector and the camera intrinsic matrix to establish the mapping relationship from the image plane to the camera coordinate system, which specifically includes: Performing tensor product operation on the global chord length feature vector and the camera intrinsic matrix, first reshape the global chord length feature vector into a two-dimensional feature matrix of 6x10 to match the structural dimension of the intrinsic matrix; then perform tensor product operation, multiply each element of the intrinsic matrix with the corresponding position element of the feature matrix to generate a high-dimensional tensor of 3x3x60, which encodes the coupling relationship between image features and camera optical characteristics; then perform dimension reduction processing on the high-dimensional tensor, extract the first 10 principal components through principal component analysis, retain more than 95% of the information amount, and obtain a 10-dimensional compressed feature vector.

[0042] Based on the 10-dimensional compressed feature vector, first perform spatial gridding processing on the compressed feature vector: decompose the 10-dimensional feature vector into 5 spatial scale factors and 5 distortion compensation factors according to its physical meaning, wherein the spatial scale factors correspond to the local scale changes of the five key regions (top left, top right, center, bottom left, and bottom right) of the image plane, and the distortion compensation factors correspond to the radial distortion and tangential distortion correction parameters of the same regions; for any point (u, v) on the image plane, determine its belonging image region, calculate the Euclidean distance between the point and the center points of the five key regions to obtain a distance weight vector, and after softmax normalization processing of the weight vector, obtain the weighted coefficients of the five spatial scale factors; then perform bilinear interpolation operation: in the u direction, calculate the interpolation weight in the u direction according to the u coordinate difference between the current point and the adjacent grid points; in the v direction, calculate the interpolation weight in the v direction according to the v coordinate difference; multiply the weights in the u direction and the v direction to obtain the comprehensive weight of the four nearest neighbor grid points, and use these weights to perform weighted average on the spatial scale factors corresponding to the four adjacent grid points to obtain the spatial scale factor S(u, v) corresponding to the point (u, v).

[0043] At the same time, pre-correct the lens distortion: use the distortion coefficients in the camera intrinsic matrix to calculate the radial distortion offset and tangential distortion offset of the point (u, v) to obtain the distortion-corrected coordinates (ud, vd); use the distortion compensation factor to perform weighted average on the (ud, vd) position according to the same spatial interpolation method to obtain the distortion compensation coefficient D(ud, vd) of the point; multiply the spatial scale factor S(u, v) and the distortion compensation coefficient D(ud, vd) to obtain the final composite scale factor K(u, v) = S(u, v) x D(ud, vd).

[0044] In constructing the mapping function, firstly, the coordinates (ud, vd) corrected by distortion are multiplied by the inverse matrix of the camera intrinsic matrix to obtain the preliminary normalized camera coordinates (Xp, Yp); then Xp is multiplied by the composite scale factor K(u, v) to obtain the final Xc coordinate, Yp is multiplied by the composite scale factor K(u, v) to obtain the final Yc coordinate, and the Zc coordinate is fixed as 1; through this spatial adaptive scale factor calculation method, each image point can obtain accurate scale correction matched with its position, especially for the image edge area, this method can effectively compensate the scale change caused by lens distortion and perspective effect, and ensure that the mapped camera coordinates (Xc, Yc, Zc) can accurately reflect the propagation direction of the actual light from the scene point to the camera optical center, thereby providing high-precision geometric basis for subsequent world coordinate system conversion.

[0045] In step 306, based on the mapping relationship from the image plane to the camera coordinate system, in combination with the ground calibration control points pre-set at the traffic light intersection, an initial perspective transformation matrix between the image plane and the world coordinate system is calculated, specifically including: Based on the mapping relationship from the image plane to the camera coordinate system established in step 305, in combination with the ground calibration control points pre-set at the traffic light intersection, firstly, the pre-set 12 calibration control points in the monitoring video stream are automatically identified: through the pre-loaded calibration control point template library, real-time feature matching is performed on the current monitoring picture, the stop line end points are identified by detecting the end point features of the white marking line and the intersection position with the lane line, the lane line intersection is positioned by detecting the straight line through Hough transform and calculating the intersection coordinates, and the crosswalk corner points are confirmed by detecting the starting end point and edge turning features of the zebra crossing; to improve the identification accuracy, sub-pixel level positioning is performed on each candidate control point, and the image coordinates of the control point are accurately determined by calculating the zero-crossing point of the image gradient field; at the same time, the system maintains a world coordinate database of the calibration control points, and these coordinates are obtained by high-precision differential GPS measurement during device installation, and the world coordinates of each control point include X, Y and Z components, wherein the Z component is uniformly set as the ground height zero point.

[0046] The image coordinates (u, v) of the 12 identified control points are converted to the camera coordinate system one by one through the mapping relationship in step 305: for each control point, the radial and tangential distortions are eliminated to obtain the corrected image coordinates; then the inverse matrix of the camera intrinsic matrix is used to convert the corrected image coordinates into normalized camera coordinates (Xc, Yc, 1); then, according to the composite scale factor K(u, v) calculated in step 305, the scale compensation is performed on the normalized coordinates to obtain the accurate three-dimensional coordinates (Xc', Yc', Zc') in the camera coordinate system.

[0047] The mapping relationship between the image coordinates and the world coordinates is established by using a direct linear transformation algorithm. First, a perspective projection equation is constructed for each calibration control point. For the i-th calibration control point, the accurate three-dimensional position in the world coordinate system is (Xwi, Ywi, Zwi), which is obtained by high-precision GPS measurement. When the point is imaged by the camera, the corresponding spatial position in the camera coordinate system is (Xa, Ya, Za), which is converted from the image coordinates by the mapping relationship in step 305. The perspective transformation matrix H is a 4x4 homogeneous transformation matrix, which transforms the homogeneous form of the world coordinate point [Xwi, Ywi, Zwi, 1] T into the homogeneous form of the camera coordinate point [Xhi, Yhi, Zhi, Whi] T .

[0048] When the equation is specifically constructed, the 4x4 perspective transformation matrix H is parameterized as 16 elements, which are recorded as h11, h12, h13, h14, h21, h22, h23, h24, h31, h32, h33, h34, h41, h42, h43, h44 in row priority order. Due to the scale invariance of the homogeneous coordinates, h44 is fixed as 1, and 15 unknown parameters are actually solved. Matrix multiplication operation is performed on the world coordinate homogeneous vector and the transformation matrix to obtain the expressions of the four homogeneous components: Xhi = h11xXwi + h12xYwi + h13xZwi + h14; Yhi = h21xXwi + h22xYwi + h23xZwi + h24; Zhi = h31xXwi + h32xYwi + h33xZwi + h34; Whi = h41xXwi + h42xYwi + h43xZwi + 1.

[0049] where (Xwi, Ywi, Zwi) represents the three-dimensional coordinates of the i-th calibration control point in the world coordinate system, and the world coordinate system is a global reference system fixed on the ground at the traffic light intersection, which is usually defined as: the X-axis is the horizontal transverse direction (perpendicular to the vehicle driving direction), the Y-axis is the horizontal longitudinal direction (along the vehicle driving direction), and the Z-axis is the vertical height direction; (Xa, Ya, Za) represents the three-dimensional coordinates of the i-th calibration control point in the camera coordinate system, and the camera coordinate system takes the camera optical center as the origin, the Z-axis is along the optical axis direction (pointing to the scene), the X-axis is horizontally to the right, and the Y-axis is vertically downward. These coordinates are converted from the image coordinates by the mapping relationship in step 305, representing the spatial position of the control point relative to the camera; [Xwi, Ywi, Zwi, 1] THomogeneous coordinate representation of world coordinates, the homogeneous term 1 is added to unify the affine transformation (including translation) as a linear transformation, T represents the transpose operation, which converts the row vector to the column vector, which is convenient for matrix multiplication operation, and the homogeneous coordinates are used in computer vision to uniformly process perspective projection, rotation, translation and other geometric transformations; [Xhi, Yhi, Zhi, Whi] T Homogeneous coordinate intermediate result after perspective transformation, which is a four-dimensional homogeneous vector obtained by transforming the world coordinates through the 4 × 4 perspective transformation matrix, which has not been converted into standard three-dimensional coordinates, wherein Whi is the homogeneous term (weight factor) for subsequent normalization processing; h11, h12, h13, h14, h21, h22, h23, h24, h31, h32, h33, h34, h41, h42, h43, h44 represent 16 elements of the 4 × 4 perspective transformation matrix H, which maps the world coordinate system to the camera coordinate system, and the meanings of the elements are as follows: h11, h12, h13, h21, h22, h23, h31, h32, h33 represent a 3 × 3 rotation submatrix, which describes the rotation relationship of the world coordinate system to the camera coordinate system; h14, h24, h34 represent the translation vector, which describes the offset of the coordinate system origin; h41, h42, h43 represent the perspective component, which describes the influence of depth information on projection; h44 represents the homogeneous scaling factor, which is usually fixed as 1 to eliminate the scale uncertainty; To convert the homogeneous coordinates into Cartesian coordinates, the first three components need to be divided by the fourth component: Xa = Xhi / Whi, Ya = Yhi / Whi, Za = Zhi / Whi. To eliminate the denominator term and linearize the equation, multiply both sides of the equation by Whi to get: Xa × Whi = Xhi, Ya × Whi = Yhi. Expand the two equations, move all terms containing unknown parameters to the left side of the equation, and move the constant term to the right side to get two linearly independent equations; each equation is a linear combination of 15 unknown parameters, and the coefficients are calculated from the known world coordinates and camera coordinates. Perform the above equation construction process for each of the 12 calibration control points to generate two linear equations for each control point, resulting in a total of 24 linear equations; arrange these equations into a standard matrix form: construct a 24 × 15 coefficient matrix A, where the k-th row corresponds to the k-th equation, and the 15 elements are the coefficients of the 15 unknown parameters in the equation; construct a 24-dimensional constant vector b, where the k-th element is the constant term of the corresponding equation. In this way, a linear equation system A × h ＝ b is formed, where h is a 15-dimensional unknown parameter vector.

[0050] Since the number of equations is greater than the number of unknown parameters, the equation set is over-determined. The singular value decomposition method is used to solve the least square solution. Before decomposition, the coefficient matrix A and the constant vector b are normalized: the mean of all the world coordinates and camera coordinates of the calibration control points is calculated, and the origin of the coordinate system is moved to the mean center; the scale factor of the coordinates is calculated, and the coordinate values are scaled to the range of [-1, 1] to improve numerical stability. Singular value decomposition is performed on the normalized coefficient matrix Am to obtain three matrices: Am = U x Σ x V T where U is a 24 x 24 orthogonal matrix, Σ is a 24 x 15 diagonal matrix (the diagonal elements are singular values arranged in descending order), and V is a 15 x 15 orthogonal matrix.

[0051] According to the least square theory, the optimal solution vector hm is equal to the last column of the V matrix (corresponding to the right singular vector of the smallest singular value), because this vector satisfies the equation constraint while minimizing ||Am x hm||. The hm is rearranged as the first 15 elements of the 4 x 4 matrix, and the 16th element is set to 1 to obtain the perspective transformation matrix hm in the normalized coordinate system. Through inverse normalization transformation, the hm is converted into the transformation matrix H in the original coordinate system: using the translation amount and weight coefficient during normalization, the corresponding elements of hm are scaled and translated to compensate, and the final 16-element perspective transformation matrix is obtained.

[0052] All 12 world coordinate points are transformed into the camera coordinate system by the matrix H, and the reprojection error of the transformation result and the actual camera coordinates is calculated; if the average error is less than 2 pixels and the maximum error is less than 5 pixels, the matrix is accepted as the initial solution; otherwise, the identification quality of the calibration control points is checked, and the abnormal points with too large error are removed to solve again. The initial perspective transformation matrix contains complete geometric transformation information, providing a high-precision initial estimate for the subsequent optimization step.

[0053] During the solving process, the RANSAC algorithm is used to enhance robustness: randomly select 6 control points to form a minimum sample set (since at least 6 points are required for perspective transformation), calculate the perspective transformation matrix under the current sample set; project the remaining 6 control points back to the image plane through the matrix, and calculate the reprojection error; set the inlier threshold to 3 pixels, i.e. the points with reprojection error less than 3 pixels are considered as inliers; record the number of inliers in the current iteration; repeat this process 1000 times, each time randomly selecting a different combination of 6 points; select the iteration result with the most inliers as the optimal model, if the maximum number of inliers is less than 8 (i.e. the inlier ratio is less than 67%), then determine that there is serious interference in the current scene, and trigger the manual review mechanism; finally, use all inliers to recalculate the perspective transformation matrix to improve accuracy.

[0054] Meanwhile, the world coordinate system is directionally constrained according to the traffic scene characteristics: taking the vehicle passing direction as the reference, the positive direction of the Y-axis of the world coordinate system is defined as the normal driving direction of the vehicle; through analyzing the spatial distribution of all the calibration control points, a scene main direction vector is calculated; the main direction vector is aligned with the preset vehicle passing direction, and the rotation component of the perspective transformation matrix is adjusted through a rotation matrix, so as to ensure that the Y-axis of the transformed world coordinate system is consistent with the actual road direction; this process is realized by minimizing the projection variance of all the control points in the Y-axis direction, so as to ensure that the physical meaning of the coordinate system meets the traffic law enforcement requirements.

[0055] The finally obtained 4×4 initial perspective transformation matrix contains complete geometric transformation information: the 3×3 sub-matrix in the upper left corner represents the rotation component, which describes the attitude change of the camera coordinate system to the world coordinate system; the first three elements in the fourth column represent the translation component, which describes the offset of the coordinate system origin; the first three elements in the fourth row represent the perspective component, which describes the scene depth information; the matrix can map any point on the image plane to the three-dimensional space position in the world coordinate system through homogeneous coordinate transformation, thereby providing a basic geometric framework for subsequent three-dimensional reconstruction.

[0056] In step 307, singular value decomposition is performed on the initial perspective transformation matrix to obtain three orthogonal matrices and a singular value vector; the distribution characteristics of the singular value vector are suppressed to obtain the orthogonal matrix after noise suppression; and based on the orthogonal matrix after noise suppression, an optimized perspective transformation matrix is reconstructed, specifically including: In step 306, singular value decomposition is performed on the initial perspective transformation matrix to obtain three matrices: U, Σ, and V T , wherein U and V T are 4×4 orthogonal matrices, and Σ is a 4×4 diagonal matrix with singular values σ1, σ2, σ3, and σ4 on the diagonal line, which are sorted in descending order and represent the stretching degree of the transformation matrix in different directions; the distribution characteristics of the singular value vector are analyzed, and the decay rate of the singular values is calculated; when the decay rate exceeds 0.5, it is considered that the subsequent singular values mainly contain noise information; the singular value vector is suppressed by setting a threshold σd=0.1×σ1, and the singular values less than the threshold are set to zero, while the corresponding column vectors of the orthogonal matrices U and V T are kept unchanged; based on the suppressed singular value diagonal matrix Σ' and the original orthogonal matrices U and V T , an optimized perspective transformation matrix H' = U×Σ'×V TDuring the reconstruction process, the matrix elements are normalized to ensure that H'[3][3] = 1, which meets the homogeneous coordinate requirement of perspective transformation; finally, the geometric consistency of the reconstructed matrix is verified, and the re-projection error of all calibration control points is calculated. If the average error exceeds 2 pixels, the suppression threshold is adjusted and the decomposition is re-executed until the matrix not only retains the main geometric information but also effectively suppresses the measurement noise and calculation error.

[0057] Step 308, the reconstructed and optimized perspective transformation matrix is fused with the geometric constraint conditions of the vehicle contour to obtain an adjusted perspective transformation matrix, specifically including: The optimized perspective transformation matrix generated by step 307 is received, and the inverse transformation relationship of the matrix is used to process a plurality of bottom key points selected from the current vehicle contour (5 points evenly distributed at the lowest part of the contour, representing the front wheel, rear wheel and chassis ground contact position). Specifically, the image pixel coordinates of each bottom key point are combined with a hypothetical initial height value (usually set to 0) representing the position on the ground to form a three-dimensional homogeneous coordinate. Then, the coordinate is multiplied by the inverse matrix of the optimized perspective transformation matrix; after multiplication, a new homogeneous coordinate is obtained, and the first three components are divided by the fourth component to solve the corresponding three-dimensional space coordinates of the contour point in the world coordinate system. Repeat this operation for all 5 bottom key points to obtain an initial set of three-dimensional space points.

[0058] Next, based on the set of three-dimensional space points, spatial plane fitting is performed. The fitting process uses the least squares principle to find a spatial plane equation that minimizes the sum of the squares of the perpendicular distances of all points to the plane. The general form of the plane equation is Ax + By + Cz + D = 0, where (x, y, z) is the coordinate of the three-dimensional space point, A, B, C and D are the coefficients of the plane equation, and (A, B, C) is the normal vector of the plane. The calculation process is as follows: first, calculate the average value of the three-dimensional coordinates of all points to obtain the center point coordinate of the point cloud; then calculate the offset of all points relative to the center and construct a covariance matrix describing the distribution of the point set. By performing eigenvalue decomposition on this covariance matrix, a set of eigenvalues and their corresponding eigenvectors are obtained. The eigenvector corresponding to the smallest eigenvalue is determined as the normal direction of the sought spatial plane. To ensure that the normal vector is a unit vector, it is usually normalized; then, using the constraint that the center point is on the plane, the value of D is calculated to determine the complete plane equation.

[0059] After obtaining the fitted plane equation, deviation analysis is performed immediately. The analysis method is: the three-dimensional coordinates obtained by back projection of each bottom key point are substituted into the fitted plane equation, and the vertical distance, i.e. height deviation, from the plane is calculated. In theory, if the perspective transformation matrix is perfect and all bottom key points are located on the same horizontal plane, the height deviation of all points should be zero. The system calculates the average value and the maximum value of the absolute values of the height deviations of all points, which are used as quantitative indicators to evaluate the consistency of the current optimized perspective transformation matrix and the coplanar physical constraint of the ground point.

[0060] Based on the results of the deviation analysis, a spatial correction transformation is constructed. The purpose of the transformation is to make the coordinates of the bottom key points closer to the plane fitted by them after a small rigid adjustment (including rotation and translation) of the world coordinate system, and to make the plane horizontal and consistent with the preset ground reference height. The specific derivation process is: first, according to the angle between the normal direction of the fitted plane and the normal direction of the ideal horizontal plane (i.e. the Z-axis direction of the world coordinate system) and the rotation axis, a correction rotation quantity is calculated, which is to rotate the fitted plane to be horizontal. Then, according to the difference between the height of the center point of the fitted plane and the preset ground reference height, a correction translation quantity is calculated, which is to adjust the plane to the correct height. Combining the rotation operation and the translation operation, a complete spatial correction transformation is constructed.

[0061] Finally, the matrix fusion operation is performed; the spatial correction transformation derived above is represented as a four-by-four homogeneous transformation matrix. Then, the correction matrix and the optimized perspective transformation matrix input in step 307 are multiplied and combined in the correct order. This multiplication order means that the original optimized matrix is used for coordinate transformation first, and then the correction transformation is applied; through this matrix multiplication operation, the mathematical relationship of the correction transformation is directly absorbed and integrated into the original perspective transformation matrix parameters, thereby generating a new adjusted perspective transformation matrix that has been fine-tuned by geometric constraints. This new matrix maintains most of the characteristics of the original mapping relationship while ensuring that the vehicle bottom ground point can meet the coplanar constraint more accurately after back projection, thereby improving the overall geometric reality of three-dimensional reconstruction.

[0062] Step 309, cascade the adjusted perspective transformation matrix and the mapping relationship from the image plane to the camera coordinate system to generate the final perspective projection model, which includes: The adjusted perspective transformation matrix of step 308 is cascaded with the mapping relationship of the image plane to the camera coordinate system established in step 305, the adjusted perspective transformation matrix He is first converted into homogeneous coordinate form to ensure its compatibility with the dimension of the camera coordinate system mapping matrix; then the cascaded transformation matrix Hf = He x Hr is calculated, wherein Hr is the mapping relationship matrix obtained in step 305, and the cascaded operation realizes the end-to-end mapping from the image plane to the world coordinate system; the condition number of the cascaded matrix is verified, the condition number cond(Hf) of the matrix is calculated, if the condition number is greater than 1000, it indicates that the matrix is ill-conditioned, and the weight ratio needs to be adjusted again; then a verification mechanism of the perspective projection model is established, 5 verification points not in the calibration control point set are selected, the re-projection error is calculated, if the average error is more than 2 pixels, it is rolled back to the last step and the constraint weight is adjusted; finally, the generated perspective projection model contains complete geometric mapping information, for any point (u, v, 1) on the image plane, the three-dimensional coordinates (Xw, Yw, Zw) in the world coordinate system can be directly calculated through Hf, the model not only considers the intrinsic characteristics of the camera, the lens distortion, but also integrates the scene calibration information and the vehicle geometric constraint, and can accurately map the vehicle contour in the two-dimensional image to the three-dimensional space, providing a high-precision projection basis for the vehicle three-dimensional reconstruction in step 4, especially when dealing with large trucks and other long-distance targets, the model can effectively eliminate the perspective distortion and ensure the accuracy of the spatial position calculation.

[0063] In a preferred embodiment of the present application, according to the perspective projection model, the spatial position data of the three-dimensional reconstructed vehicle is determined, comprising: Step 310, the perspective projection model is registered and aligned with the preset vehicle priori geometric model, which contains the length-width-height ratio constraint and the contour topology structure of the standard vehicle, specifically including: the final perspective projection model generated in step 309 is registered and aligned with the preset vehicle priori geometric model, first load the vehicle priori geometric model library, which contains the standardized three-dimensional geometric model for different vehicle types (sedan, SUV, truck, bus), each model is composed of 500-2000 triangular patches, which accurately represents the appearance contour of the vehicle; the model contains strict length-width-height ratio constraint, such as the length-width-height ratio range of sedan is 4.5:1.8:1.4, and the length-width-height ratio of truck is 8:2.5:3.2, and the key contour topology structure is defined, including the curvature continuity constraint of the front face of the vehicle head, the parallel line constraint of the side of the vehicle body, and the rectangular topology constraint of the vehicle window area; the registration and alignment process adopts a hierarchical matching strategy: first, based on the global chord length feature vector obtained in step 303, the closest vehicle type model is automatically selected by feature similarity matching; then project the selected vehicle priori model to the image plane, and calculate its expected contour in the two-dimensional image by using the perspective projection model; the initial alignment quality is evaluated by calculating the Hausdorff distance between the expected contour and the actual vehicle contour extracted in step 300; if the distance exceeds the preset threshold (usually 10 pixels), adjust the initial pose of the model, and perform rough translation along the X, Y and Z directions of the world coordinate system, so that the model center and the vehicle image center are roughly coincident; at the same time, according to the major axis direction of the ellipse fitted in step 301, the rotation angle of the model is initialized, so that the longitudinal axis of the model is consistent with the main direction of the vehicle in the image; during the registration process, special attention is paid to the key geometric feature points of the vehicle, such as the front end point of the vehicle head, the rear end point of the vehicle tail, and the highest point of the vehicle roof, etc., and through local feature matching, these key points are accurately aligned with the corresponding points in the image after projection; finally output the preliminary registered vehicle three-dimensional model, which has a reasonable initial position and orientation in the world coordinate system.

[0064] Step 311, the pose parameters of the vehicle in the world coordinate system are calculated by optimizing the matching error between the vehicle prior geometric model and the perspective projection model through the iterative closest point algorithm, and the pose parameters include the translation vector and the rotation matrix, specifically including: based on the vehicle three-dimensional model preliminarily registered in step 310, the matching error between the vehicle prior geometric model and the perspective projection model is optimized through the iterative closest point algorithm, first 2000 three-dimensional points are uniformly sampled from the surface of the vehicle prior geometric model to form a model point set; At the same time, 1500 edge points are extracted from the actual vehicle contour extracted in step 300, and these two-dimensional image points are back projected to the three-dimensional space through the perspective projection model to form a scene point set; The iterative closest point algorithm executes the following loop process: in each iteration, for each point in the model point set, find the closest corresponding point in the scene point set, and calculate the Euclidean distance between the two as the matching error; The RANSAC algorithm is used to remove abnormal corresponding point pairs, and the distance threshold is set to 0.5 meters. Points exceeding the threshold are considered as outliers; Based on the correspondence of the inner points, a least squares optimization problem is constructed, and the goal is to minimize the sum of the squares of the distances of all inner point corresponding pairs; The optimal rigid transformation (including the rotation matrix R and the translation vector t) is solved by the singular value decomposition method, so that the matching error of the transformed model point set and the scene point set is minimized; The transformation parameters obtained are applied to update the pose of the vehicle model, and the matching error is recalculated; Set the convergence condition: when the average matching error reduction of three consecutive iterations is less than 0.01 meters, or the number of iterations reaches 50, terminate the optimization process; In the optimization process, the kinematic constraint of the vehicle is introduced: according to the continuity of the vehicle motion between consecutive frames, the change amplitude of the translation vector in each iteration is limited to not more than 0.2 meters, and the change amplitude of the rotation angle is limited to not more than 5 degrees, avoiding the jump of the optimization result; At the same time, combined with the geometric constraint of the traffic scene, the Z coordinate of the vehicle bottom point set in the world coordinate system is forced to be close to the ground height (usually 0.1-0.3 meters), ensuring that the vehicle attitude meets the actual physical conditions; Finally, the optimized pose parameters are output, including a 3x3 rotation matrix R (describing the orientation of the vehicle in the world coordinate system) and a 3x1 translation vector t (describing the position coordinates of the vehicle center in the world coordinate system), which accurately represent the pose state of the vehicle in three-dimensional space.

[0065] At step 312, based on the pose parameters, the key feature points on the vehicle contour are subjected to three-dimensional coordinate back-projection to generate three-dimensional point cloud data of the vehicle in the monitoring area, specifically including: based on the pose parameters (rotation matrix R and translation vector t) obtained by optimization at step 309, the key feature points on the vehicle contour are subjected to three-dimensional coordinate back-projection, first, 50 key feature points in the vehicle contour are identified, including: the most front end point of the vehicle head, the last end point of the vehicle tail, the left and right side outermost points, the highest point of the vehicle roof, the four wheel grounding points, the vehicle window corner points, the vehicle door handle positions, etc., which have a clear three-dimensional coordinate in the vehicle prior geometric model; the model coordinates of each key feature point are converted to the world coordinate system through the pose parameters: first, the direction is adjusted by applying the rotation matrix R, and then the position is offset by applying the translation vector t, to obtain the accurate three-dimensional coordinates of each key point in the world coordinate system; at the same time, the inverse transformation of the perspective projection model is used to back-project the corresponding feature point coordinates on the image plane to the three-dimensional space to generate auxiliary three-dimensional point estimates; through a weighted fusion strategy, the model-driven three-dimensional coordinates and the image-driven three-dimensional coordinates are fused: for the clear contour area (such as the vehicle head and tail), the image-driven coordinates are given a higher weight (0.7); for the complex texture area (such as the vehicle window and door), the model-driven coordinates are given a higher weight (0.8); the fused three-dimensional point coordinates are subjected to outlier detection processing, the average distance of each point from the neighboring points is calculated, and the points with abnormal distance (points exceeding 1.5 times the standard deviation of the average distance) are removed; to improve the point cloud density, interpolation processing is performed between the key feature points: along the main direction of the vehicle contour, 5 intermediate points are uniformly inserted between adjacent key points, and the three-dimensional coordinates of these intermediate points are calculated through spline curve interpolation; at the same time, 10 layers are uniformly sampled from the bottom to the top in the vehicle height direction, and the contour points of the vehicle cross section are generated in each layer; finally, a sparse point cloud data containing 2000 three-dimensional points is generated, which are accurately distributed on the vehicle surface, covering various key parts of the vehicle, and each point is given a confidence weight reflecting its geometric reliability; the point cloud data is subjected to smoothing filter processing, the moving least squares method is used to fit the surface of the local area, and the surface discontinuity caused by noise and discretization is eliminated, and a high-quality three-dimensional point cloud representation of the vehicle is output.

[0066] Step 313, convert the three-dimensional point cloud data into a spatial occupancy grid, calculate the vehicle center point coordinates, spatial orientation angle and circumscribed cube size, form a three-dimensional spatial position data set containing the vehicle accurate spatial position, motion direction and physical size, specifically including: converting the three-dimensional point cloud data into a spatial occupancy grid, first defining a three-dimensional grid space surrounding the entire vehicle, the grid resolution is set to 0.1m x 0.1m x 0.1m, ensuring that the details of the vehicle can be captured; map each three-dimensional point to the corresponding grid cell, and count the point cloud density contained in each grid cell; through connected component analysis, identify the continuous grid area occupied by the vehicle point cloud, and remove the isolated noise grid; based on the occupancy grid, calculate the geometric center point coordinates of the vehicle: weighted average the center coordinates of all occupied grids, and the weight is the point cloud density in the grid, to obtain the accurate center position (Xr, Yr, Zr) of the vehicle in the world coordinate system; calculate the spatial orientation angle of the vehicle: extract the main direction vector of the vehicle bottom contour, determine the main axis direction of the contour point set through principal component analysis, calculate the included angle between the direction vector and the Y axis (vehicle driving direction) of the world coordinate system, and obtain the yaw angle ψ of the vehicle; at the same time, calculate the pitch angle θ and the roll angle φ of the vehicle, which are obtained by analyzing the height difference distribution of the vehicle top point set and the bottom point set; construct the circumscribed cube of the vehicle: on the basis of the occupancy grid, along the three main axes of the vehicle local coordinate system, calculate the maximum and minimum projection values of the point cloud in each direction, to obtain the eight vertex coordinates of the circumscribed cube; the size parameters of the circumscribed cube include length L (along the vehicle longitudinal direction), width W (along the vehicle transverse direction), and height H (along the vehicle vertical direction), which are optimized according to the statistical characteristics of the point cloud distribution, excluding the influence of outliers; in order to improve the size accuracy, the prior knowledge of the vehicle type is introduced: for vehicles identified as trucks, the height H is constrained to be not less than 2.5 meters; for cars, the length L is constrained to be within the range of 4-5 meters; integrate the calculated center point coordinates, spatial orientation angle (ψ, θ, φ), circumscribed cube size (L, W, H) and point cloud density distribution characteristics into a structured three-dimensional spatial position data set; the data set contains timestamp information, recording the accurate time when the data is generated; at the same time, it contains a confidence index, which evaluates the reliability of three-dimensional reconstruction based on point cloud density, matching error, geometric consistency and other multidimensional; the finally output three-dimensional spatial position data set provides an accurate spatial reference for subsequent vehicle motion trajectory prediction, illegal judgment and snapshot timing control, especially in complex traffic scenes, it can accurately distinguish the spatial position relationship of adjacent vehicles, and avoid misjudgment.

[0067] In a preferred embodiment of the present application, step 4, based on the vehicle spatial position data of three-dimensional reconstruction, a pyramid geometric model of the vehicle occupancy space is constructed, including: In step 401, the spatial position of the pyramid vertex is determined along the motion direction extension line according to the motion direction vector and the current speed parameter of the vehicle, the pyramid vertex being located at a preset safety distance in front of the vehicle, the preset safety distance being dynamically adjusted according to the vehicle type and the speed, specifically including: Based on the three-dimensional spatial position data set generated in step 311, the motion direction vector and the current speed parameter of the vehicle are first extracted. The motion direction vector is calculated by the center point coordinates of three consecutive frames: taking the center point coordinates of the current frame, the previous frame and the frame two frames ago, two displacement vectors are constructed, the instantaneous jitter is eliminated by vector averaging, and a smooth motion direction unit vector is obtained; the current speed parameter is obtained by calculating the Euclidean distance between the center points of adjacent frames divided by the time interval (usually 1 / 25 second), and is smoothed by a Kalman filter; according to the vehicle type (obtained from the vehicle prior geometric model matching result in step 308) and the current speed, the preset safety distance is dynamically calculated: for a small car, the basic safety distance is 5 meters, and for each increase of 10 km / h speed, 1 meter is added; for a large truck, the basic safety distance is 10 meters, and for each increase of 10 km / h speed, 2 meters are added; for emergency braking state (detected by acceleration parameter, when the deceleration is more than 3 m / s 2 2), the safety distance is additionally increased by 30%. Based on the center point coordinates of the vehicle, the initial spatial position of the pyramid vertex is obtained by extending the preset safety distance along the motion direction vector; in order to adapt to the road curvature, the position is corrected by road geometry: the high-precision map data of the traffic light intersection is obtained, the curvature information of the lane center line is extracted, the pyramid vertex is projected onto the nearest lane center line, and the vertex position is ensured to conform to the actual road direction; at the same time, the influence of slope is considered, the Z coordinate of the vertex is adjusted according to the digital elevation model of the intersection, so that it is consistent with the road slope; the finally determined pyramid vertex coordinates (xv, yv, zv) are located at a safety distance in front of the vehicle, accurately representing the farthest boundary of the front safety buffer area of the vehicle in the current motion state.

[0068] Step 402, project the key feature points on the vehicle contour to the ground plane of the traffic light intersection to form a vehicle bottom contour polygon; perform convex hull processing on the vehicle bottom contour polygon to generate a regular polygon of the pyramid bottom surface, specifically including: based on the vehicle three-dimensional point cloud data generated in step 310, extracting 80 key feature points on the vehicle contour, which are uniformly distributed on the vehicle bottom edge, including the front bumper lower edge, the rear bumper lower edge, the left and right side skirts, the wheel grounding point and the like; project these key feature points to the ground plane of the traffic light intersection: the ground plane is determined by the ground control points pre-calibrated, and its equation is Z = 0 (the Z axis is vertically upward in the world coordinate system, and the ground is the XY plane); for each key feature point (xi, yi, zi), keep the X and Y coordinates unchanged, and set the Z coordinate to 0.15 meters (considering the vehicle ground clearance), to obtain the projection point (xi, yi, 0.15), forming a vehicle bottom contour point set. Perform cluster analysis on the bottom contour point set, and use the DBSCAN algorithm to identify the main contour connected region, and remove outliers caused by noise or occlusion; arrange the remaining effective projection points in clockwise order to construct an initial vehicle bottom contour polygon; perform convex hull processing on the polygon: using the Graham scan algorithm, first find the left lower corner point as the reference point, calculate the polar angle of other points relative to the reference point and sort, then check the turning (left or right) of each three consecutive points in turn, remove the middle point causing right turning, and finally obtain a convex polygon containing 12-16 vertices; to adapt to the actual shape of the vehicle, locally optimize the convex hull result: in the vehicle head region, adjust the position of the front vertex according to the elliptical fitting result of step 301 to make it more fit the actual vehicle head contour; in the vehicle tail region, according to the geometric features of the tail lamp and rear bumper, fine-tune the distribution density of the rear vertices; the finally generated regular polygon of the pyramid bottom surface has a smooth geometric shape, accurately reflects the projection occupation range of the vehicle on the ground, and provides an accurate bottom surface boundary for subsequent pyramid construction.

[0069] Step 403, connecting the pyramid vertex and each vertex of the regular polygon of the pyramid base, constructing a closed pyramid geometric model representing the three-dimensional space range occupied by the vehicle in the monitoring area, including the space volume of the vehicle body and the front safety buffer area, specifically including: based on the pyramid vertex (xv, yv, zv) determined in step 401 and the regular polygon (including N vertices P1, P2,..., PN) of the pyramid base generated in step 402, constructing a closed pyramid geometric model. First, establish the topology of the pyramid: connect the pyramid vertex with each vertex of the base polygon to form N triangular sides, each side being composed of vertex V and two adjacent vertices Pi and Pi+1 of the base (PN+1=P1); at the same time, the base polygon itself serves as the base of the pyramid, composed of N triangular facets (the polygon is decomposed into triangles by a triangulation algorithm). Geometric optimization of the pyramid model: calculate the normal vector of each side triangle to ensure that all normal vectors point outward, ensuring the closure and consistency of the model; smooth the base polygon, insert transition triangles between adjacent sides to eliminate the corner effect caused by the limited number of base vertices, making the pyramid surface more continuous. To accurately represent the space volume occupied by the vehicle, physically constrain the pyramid model: in the vehicle head area, according to the actual length of the vehicle, offset the pyramid vertex by 20% towards the vehicle center to avoid excessive extension of the safety buffer area; in the side area, according to the width variation characteristics of the vehicle, add control points at the vehicle door position to make the pyramid side surface conform to the actual contour of the vehicle; in the height direction, set different height constraints according to the vehicle type: the height of a sedan is limited to 1.8 meters, the height of an SUV is 2.2 meters, and the height of a truck is 3.5 meters, to ensure that the pyramid model does not exceed the actual physical height of the vehicle. The final pyramid geometric model contains complete three-dimensional space information: the surface is composed of M triangular facets (usually M=2N+10), the volume is calculated by integrating the triangular facets, and the centroid position is determined by the geometric center algorithm; this model not only contains the space range of the vehicle body, but also contains the front safety buffer area, and can dynamically adapt to the motion state of the vehicle and the road conditions, providing an accurate geometric basis for subsequent space volume calculation and violation judgment.

[0070] In a preferred embodiment of the present application, the spatial volume distribution of the vehicle in the monitoring area is calculated according to the pyramid geometric model, and a three-dimensional space state parameter set containing the vehicle spatial pose, motion vector and volume characteristics is generated, including: At step 404, the pyramid geometry model is spatially divided, and the pyramid is equally divided into multiple horizontal slice layers along the height direction. Specifically, based on the pyramid geometry model constructed at step 403, the height range of the pyramid is first determined: the Z coordinates of the pyramid vertex and all vertices of the bottom polygon are extracted, the maximum Z value (vertex height) and the minimum Z value (bottom surface height Zmin) are calculated, and the total height of the pyramid Ha is obtained. The number of horizontal slice layers Ns is dynamically determined according to the vehicle type and the accuracy requirement: for small vehicles (sedans, SUVs), Ns = 15 layers is set; for large vehicles (trucks, buses), Ns = 25 layers is set, ensuring that the thickness of each layer does not exceed 0.2 meters; the thickness of each layer is calculated as Ah = Ha / Ns; the horizontal slice layers are constructed from the bottom surface along the height direction: the Z coordinate range of the k-th layer (k = 0, 1,..., Ns-1) is [Zmin+k Ah, Zmin+(k+1) Ah]; for each slice layer, the intersection line with the pyramid side surface is calculated: the intersection operation is performed between each triangular side surface of the pyramid and the horizontal plane (Z = Zk) of the current slice layer to obtain the cross-section contour polygon of the layer; the topological verification is performed on the cross-section contour polygon to ensure that it is a simple polygon (without self-intersection), and if a complex polygon is detected, the polygon Boolean operation is used to decompose it into multiple simple polygons; to improve the calculation efficiency, the cross-section contours of adjacent slice layers are spatially related: the contour vertices of the k-th layer are used as the initial vertex set of the k+1-th layer, and local optimization is performed to reduce repeated calculation; finally, Ns horizontal slice layers are generated, each slice layer contains one or more cross-section contour polygons and the corresponding Z coordinate range, and these slice layers completely represent the spatial distribution characteristics of the pyramid in the height direction, providing a hierarchical geometric basis for subsequent meshing processing.

[0071] Step 405, grid processing is performed on each horizontal slice layer, and the spatial occupancy probability density of each grid cell is calculated, which is weighted and distributed according to the coincidence degree of the grid cell and the actual vehicle contour, specifically including: grid processing is performed on each horizontal slice layer generated in step 404, first, the grid resolution is determined according to the cross-sectional contour range of the slice layer: in the X-Y plane, the bounding box (minimum circumscribed rectangle) of the cross-sectional contour is calculated, and the grid size is adaptively set according to the contour size, ensuring that the length of each grid cell is between 0.05-0.15 meters, which can ensure accuracy and avoid calculation redundancy; the bounding box is divided into a uniform grid array, each grid cell has a unique two-dimensional index (i, j) and a corresponding quadrilateral area; the spatial occupancy probability density of each grid cell is calculated: first, the coincidence degree of the grid cell and the cross-sectional contour is judged, and the ray intersection method is used to accurately calculate the area ratio ar of the grid cell covered by the contour; at the same time, the actual physical characteristics of the vehicle are considered, the point cloud density information corresponding to the current height layer is extracted from the three-dimensional point cloud data in step 310, and the point cloud density is normalized to a weight wy in the range of [0, 1]; the vehicle motion state weight is introduced, and the dynamic influence coefficient wg of the current slice layer is calculated according to the motion vector in step 311 (the weight of the front region is higher, and the weight of the rear region is lower); comprehensive consideration of the above factors, the spatial occupancy probability density P(i, j, k) = ar x (0.6 x wy + 0.4 x wg) is calculated, where k represents the slice layer index; special processing is performed on the boundary grid cells: when the grid cell is partially located inside the contour, sub-grid sampling technology is used to randomly generate 100 sampling points inside the grid, and the proportion of points located inside the contour is taken as the accurate ar; in order to eliminate the discretization error, the probability density between adjacent grid cells is smoothed: the P value of each grid cell and the P value of its 8-neighbor grid cells are weighted and averaged, and the weight decays with distance, ensuring the continuity of the probability density field; finally, a two-dimensional probability density matrix is generated for each slice layer, which completely describes the distribution characteristics of the spatial occupancy in the height layer and provides an accurate probabilistic representation for volume feature calculation.

[0072] At step 406, based on the spatial occupancy probability density of each grid cell, the overall volume distribution characteristics of the pyramid geometric model are calculated, including the volume barycenter coordinates, volume distribution variance and spatial occupancy rate, specifically including: based on the spatial occupancy probability density of each grid cell calculated at step 405, the overall volume distribution characteristics of the pyramid geometric model are calculated. First, the volume barycenter coordinates are calculated: each grid cell is regarded as a mass body, and the mass thereof is equal to the probability density P(i, j, k) multiplied by the grid volume (△x×△y×△h); the weighted centroid of all grid cells is calculated, X barycenter =∑(P(i, j, k)×xi×Vcell) / ∑(P(i, j, k)×Vcell), and Y barycenter and Z barycenter are calculated in the same way, wherein xi is the X coordinate of the grid center, and Vcell is the volume of a single grid; to improve the accuracy, an adaptive integration strategy is adopted in key areas such as the front and rear of the vehicle, the grids in these areas are further subdivided, the local barycenter is recalculated, and then weighted fusion is performed. Then, the volume distribution variance is calculated: taking the volume barycenter as the reference point, the Euclidean distance d(i, j, k) of each grid cell to the barycenter is calculated; the weighted variance Var =∑[P(i, j, k)×(d(i, j, k)-μ) 2 ] / ∑P(i, j, k) is calculated, wherein μ is the average distance; the component variances in X, Y and Z directions are calculated respectively, representing the dispersion degree of the volume in each dimension; the skewness and kurtosis of the volume distribution are also calculated, reflecting the symmetry and sharpness of the distribution. Finally, the spatial occupancy rate is calculated: the number Np of grid cells with a probability density greater than 0.5 is counted, and then the total spatial occupancy rate is obtained by dividing the total number Ntotal of grid cells; the spatial occupancy rate is calculated layer by layer to analyze the occupancy characteristics of different height layers; the effective volume Vf =∑[P(i, j, k)×Vcell] is calculated, and compared with the theoretical volume of the pyramid to obtain the volume utilization rate; these volume distribution characteristics are integrated into structured data: including the volume barycenter coordinates (Xg, Yg, Zg), the overall variance σa, the component variances (σx, σy, σz), the spatial occupancy rate ρy, the effective volume Veff and other parameters; through outlier detection, the rationality of the calculation result is verified, such as when the volume barycenter deviates from the geometric center of the vehicle by more than 1 meter, the probability density calculation process is rechecked; the finally output volume distribution characteristics accurately quantify the occupancy characteristics of the vehicle in the three-dimensional space, and provide key spatial state information for the dynamic snapshot strategy.

[0073] Step 407, spatiotemporal fusion of volume distribution features with three-dimensional spatial position data set to generate three-dimensional spatial state parameter set containing vehicle real-time spatial posture angle, three-dimensional motion vector, volume distribution feature and spatial occupancy state, specifically including: spatiotemporal fusion of volume distribution features calculated in step 406 with three-dimensional spatial position data set generated in step 311, first establishing a unified time reference: taking the timestamp tu of the current frame as the reference, extracting the three-dimensional spatial position data of the previous 5 frames and the next 5 frames from the historical data buffer to form an 11-frame time window; time alignment of the volume distribution features of each historical frame, compensating for the time delay through linear interpolation to ensure that all features are under the same time reference; building a multi-dimensional feature fusion framework: calculating the difference between the volume center of gravity coordinates and the vehicle center point coordinates to obtain the volume offset vector; correlating the volume distribution variance with the vehicle motion vector and comparing the spatial occupancy rate with the vehicle size parameters to verify the abnormal state (such as a spatial occupancy rate below 0.7, which may indicate partial occlusion of the vehicle). Real-time spatial posture angle fusion: consistency verification of the orientation angle (ψ, θ, φ) of step 311 with the main direction of the volume distribution, when the angle difference exceeds 5 degrees, weighted average is used for correction, and the weight is based on the confidence of the volume distribution; three-dimensional motion vector enhancement: on the basis of the original motion vector, the volume change rate is added as an acceleration correction factor, when the volume distribution variance increases rapidly, it is determined that the vehicle is in an acceleration state, and the motion vector is adjusted accordingly; volume distribution feature integration: after standardizing the volume center of gravity, variance, occupancy rate and other parameters, they are spliced with the spatial position and motion state parameters to form a high-dimensional feature vector; spatial occupancy state generation: based on the grid probability density field, an octree spatial index structure is constructed to support fast spatial query operations such as "determining whether a point is within the vehicle occupancy space" and "calculating the spatial overlap rate of two vehicles". The finally generated three-dimensional spatial state parameter set contains: spatial posture (position coordinates, orientation angle, posture confidence), motion state (velocity vector, acceleration vector, motion trend prediction), volume feature (volume center of gravity, distribution variance, occupancy rate, effective volume), spatial occupancy (octree index, occupancy probability field, spatial relationship matrix); the parameter set has time continuity, and the state is smoothly transitioned when each frame is updated.

[0074] In a preferred embodiment of the present application, step 5, according to the three-dimensional spatial state parameter set, the coordinates of the vertices of the vehicle motion path are calculated, the final imaging time of the vehicle at the key monitoring point is determined, and the final snapshot control instruction sequence containing the front, middle and rear of the vehicle is generated, including: Step 500, extract the three-dimensional motion vector, spatial attitude angle and historical trajectory point sequence of the vehicle from the three-dimensional space state parameter set, specifically including: extracting the motion feature data of the vehicle from the three-dimensional space state parameter set generated in step 407, first accessing the motion state module in the parameter set to obtain the three-dimensional motion vector of the current frame, which contains the velocity components (Vx, Vy, Vz) in X, Y and Z directions, wherein Vx represents the lateral velocity, Vy represents the longitudinal velocity (along the road direction), and Vz represents the vertical velocity; At the same time, the acceleration vector (Ax, Ay, Az) in the motion module is extracted for subsequent trajectory prediction correction. Then access the spatial attitude and extract the spatial attitude angle data of the vehicle, including the yaw angle ψ (horizontal rotation angle around the Z axis), the pitch angle θ (forward and backward inclination angle around the X axis) and the roll angle φ (lateral inclination angle around the Y axis), which are all processed by low-pass filtering to eliminate high-frequency jitter. Then extract the historical trajectory point sequence of the last 15 frames from the historical trajectory buffer, which contains the space-time coordinates of the vehicle center point in the world coordinate system: (Xt-14, Yt-14, Zt-14, t-14), (Xt-13, Yt-13, Zt-13, t-13),..., (Xt, Yt, Zt, t), where t represents the current timestamp; Quality screening is performed on the historical trajectory points: calculate the distance change rate between adjacent points, and remove abnormal points with a distance mutation of more than 2 meters; At the same time, check the continuity of the time interval and linearly interpolate the missing frames for compensation; In order to adapt to the motion characteristics of different vehicles, the weight distribution of the historical trajectory points is dynamically adjusted according to the vehicle type identifier (sedan, SUV, truck) in the parameter set, and large vehicles are given a longer historical window weight; Finally, output the structured motion feature data packet, which contains the smoothed three-dimensional motion vector, accurate spatial attitude angle, purified historical trajectory point sequence, and corresponding timestamp sequence, providing high-quality input data for trajectory modeling.

[0075] Step 501, based on the historical trajectory point sequence and the three-dimensional motion vector, calculate the parabolic equation parameters of the vehicle motion trajectory, including the quadratic term coefficient, the linear term coefficient and the constant term, specifically including: based on the historical trajectory point sequence and the three-dimensional motion vector extracted in step 500, construct a parabolic mathematical model of the vehicle motion trajectory. First, project the historical trajectory points onto the road plane (XY plane) and ignore the Z-axis height change to simplify the calculation; Select the trajectory modeling dimension according to the vehicle type: for straight vehicles, use a two-dimensional parabolic model; For turning vehicles, use a segmented parabolic model. In two-dimensional modeling, take the Y coordinate (longitudinal position) of the historical trajectory point as the independent variable and the X coordinate (lateral position) as the dependent variable to establish a quadratic function relationship X=aY 2 +bY+c; Use weighted least squares method for parameter estimation: assign higher weight to recent trajectory points (weight coefficient=0.9 n, n is the frame number), to ensure that the model is more sensitive to the current motion state; the three-dimensional motion vector is used as a constraint condition to calculate the tangent slope dX / dY = (Vx / Vy) of the current point, which forces the derivative of the parabola at the current point to equal the slope value; the Lagrange multiplier method is used to solve the optimization problem with constraints to obtain the quadratic term coefficient a, the linear term coefficient b, and the constant term c; for a turning vehicle, a segmented modeling strategy is adopted: the change rate of the spatial attitude angle ψ is used to detect the turning starting point, and the trajectory is divided into a straight section and a turning section, and different parabolic equations are fitted respectively; in the turning section, the road curvature radius is introduced as an additional constraint to ensure that the fitted trajectory is consistent with the road geometric characteristics; residual analysis is performed on the fitting results, the distance from each historical point to the parabola is calculated, and if the average residual exceeds 0.3 meters, the model complexity is increased or the weight distribution is adjusted; finally, the optimized parabola equation parameter set is output, including the coefficients a, b, and c, and the fitting confidence index, which accurately describes the vehicle's motion trajectory trend on the road plane and provides a mathematical basis for key point calculation.

[0076] Step 502, according to the parabola equation parameters, the vertex coordinates of the parabola trajectory are calculated, which correspond to the highest or lowest point of the vehicle motion trajectory, representing the key turning position of the vehicle motion, specifically including: according to the parabola equation parameters calculated in step 501, the vertex coordinates of the parabola trajectory are calculated. For the standard parabola equation X = aY 2 + bY + c, the longitudinal coordinate Yr of the vertex is calculated by the formula Yr = -b / (2a), which corresponds to the extreme point of the parabola; Yr is substituted into the parabola equation to calculate the corresponding transverse coordinate Xr = a(Yr) 2+ b(Yr) + c; the vertical coordinate Zr of the vertex is determined according to road slope information, the slope angle a of the current road section is extracted from the high-precision map, and Zr = Zn + (Yr-Yn) tan(a) is calculated, wherein Zn is the current point height; in a physical sense, the vertex coordinate represents a key turning position of the vehicle motion trajectory: when a > 0, the vertex is the lowest point of the trajectory, which usually corresponds to the bottom of a downhill road section; when a < 0, the vertex is the highest point of the trajectory, which usually corresponds to the top of an uphill road section or an extreme point on the outside of a turning arc; in order to adapt to complex road scenes, the vertex coordinate is corrected according to road constraints: the calculated vertex is projected onto the nearest lane center line to ensure that it conforms to the actual road direction; at the same time, it is checked whether the vertex is located within the monitoring area, and if the vertex exceeds the effective monitoring range (more than 50 meters away from the camera), the boundary of the monitoring area is used as a substitute vertex; according to the principle of energy conservation, if the vertex is the highest point, the speed should be slightly lower than the current speed; if it is the lowest point, the speed should be slightly higher than the current speed, and if the deviation exceeds 15%, the parameter calculation is rechecked; finally, the three-dimensional vertex coordinates (Xr, Yr, Zr) are output, as well as the corresponding physical meaning label (highest point / lowest point) and confidence score, which serves as a key reference point for trajectory analysis and provides a reference position for subsequent snapshot point positioning.

[0077] Step 503, according to the stop line position of the traffic light intersection and the vehicle type parameter, determine the spatial coordinates of the vehicle head snapshot point, the vehicle body middle snapshot point and the vehicle tail snapshot point on the parabolic trajectory, specifically including: based on the vertex coordinate calculated in step 502 and the three-dimensional space state parameter in step 407, combined with the geometric information of the traffic light intersection, determine the spatial coordinates of the three key snapshot points. First, obtain the accurate position coordinates of the stop line from the intersection high-precision map, including the stop line center point coordinates (Xstop, Ystop, Zstop) and the normal vector direction; dynamically adjust the snapshot point offset distance according to the vehicle type parameter (extracted from the three-dimensional space state parameter set): for small cars, the vehicle head snapshot point is 3.5 meters away from the stop line; for large trucks, the vehicle head snapshot point is 6.0 meters away from the stop line; the vehicle body middle snapshot point is located L / 2 behind the vehicle head snapshot point, wherein L is the vehicle length (obtained from the circumscribed cube size in step 311); the vehicle tail snapshot point is located L behind the vehicle head snapshot point. Position the snapshot points on the parabolic trajectory: take the stop line position as the reference, measure the preset distance along the parabolic trajectory in the forward direction (vehicle driving direction) to obtain the longitudinal coordinate Yhead = Ystop + offsethead of the vehicle head snapshot point, wherein offsethead represents the longitudinal offset distance of the vehicle head snapshot point relative to the stop line in the vehicle driving direction; substitute Yhead into the parabolic equation X = aY 2+bY+c, calculate the corresponding lateral coordinate Xhead; the vertical coordinate Zhead is calculated based on the road slope, Zhead = Zstop + (Yhead - Ystop) · tan(α). For the capture point at the middle of the vehicle, Ymid = Yhead - L / 2, and for the capture point at the rear of the vehicle, Ytail = Yhead - L, calculate the corresponding Xmid, Zmid, Xtail, and Ztail respectively. To adapt to turning scenarios, curvature compensation is applied to the lateral coordinates: calculate the radius of curvature R of the parabola at the capture point, and adjust the lateral offset ΔX = W based on the vehicle width W. 2 / (8R) ensures the capture point is located on the vehicle's center trajectory. Simultaneously, considering camera view constraints, it checks if the capture point is within the camera's effective field of view (horizontal field of view ±30 degrees, vertical field of view ±15 degrees). If it exceeds this range, the capture point position is finely adjusted forward and backward along the trajectory to ensure image quality. Finally, the precise three-dimensional coordinates of three capture points are output: the front capture point (Xhead, Yhead, Zhead), the middle capture point (Xmid, Ymid, Zmid), and the rear capture point (Xtail, Ytail, Ztail). Each point contains position coordinates, expected attitude angle, and predicted image quality value, providing a spatial reference for time prediction.

[0078] Step 504: Based on the vehicle's current speed, acceleration, and spatial attitude angle, calculate the predicted arrival time of the vehicle at each capture point, and generate capture timing control parameters including timestamps, specifically including: The core motion data of the vehicle at present is extracted from the three-dimensional spatial state parameter set, including the three-dimensional velocity vector smoothed by Kalman filtering (focusing on the longitudinal velocity along the road travel direction, while referring to the lateral and vertical velocities for correction), the acceleration vector (including longitudinal acceleration and deceleration and lateral acceleration, used to determine whether the vehicle is in an accelerating, decelerating or constant speed state), and the spatial attitude angles (yaw angle, pitch angle, roll angle) processed by low-pass filtering; combined with the three-dimensional coordinates of the three capture points determined in step 503, the spatial straight-line distance from the current center point of the vehicle to each capture point is calculated, and then the spatial distance is corrected to the effective distance on the actual driving path of the vehicle according to the actual road direction (extracting lane centerline information from high-precision map) (eliminating the error caused by lateral offset and ensuring that the distance calculation fits the actual driving trajectory).

[0079] Calculate the predicted time based on the vehicle's current motion mode (constant speed, acceleration, deceleration, turning): If the vehicle is in a uniform state (the absolute value of acceleration is less than a preset threshold, usually 0.2 m / s²), directly divide the effective distance by the current longitudinal speed to get the basic time prediction value; if the vehicle is in an acceleration or deceleration state, adjust the time calculation according to the acceleration size and direction: when accelerating, the speed of the vehicle reaching the snapshot point will be higher than the current speed, and the effective driving time needs to be corrected according to the acceleration trend (when accelerating, the time is shortened, and when decelerating, the time is lengthened); if the vehicle is in a turning state (the rate of change of yaw angle is greater than a preset threshold, usually 3 degrees per frame), refer to the road curvature radius and vehicle type characteristics (when large vehicles turn, the speed usually decreases by 10%-20%), first correct the current speed (when turning, the speed is lowered), and then calculate the time combining the corrected speed and the effective distance to avoid time prediction deviation caused by turning speed change.

[0080] Extract the vehicle motion parameters (speed, acceleration, displacement) of the last 10 frames, calculate the stability coefficient of the motion state, and if the stability is high (fluctuation is less than 5%), the basic time prediction value is used as the main value; if the fluctuation is large (such as sudden acceleration or deceleration), the weighted average method is used to fuse the historical time prediction trend (recent frames have higher weight, and distant frames have lower weight) to get the smoothed time prediction value; assign a unique timestamp identifier to each snapshot point (vehicle head, middle of vehicle body, vehicle tail), and the timestamp is accurate to the millisecond level (based on the system unified time reference); the timing control parameters also include auxiliary information: time prediction confidence of each snapshot point (comprehensively evaluated according to motion state stability, distance, and road environment complexity, confidence range 0-1, closer to 1 means more reliable prediction), time adjustment threshold (maximum time range allowed for subsequent dynamic correction, usually ±50 milliseconds), and snapshot priority (vehicle head snapshot priority is the highest, ensuring that the core evidence is acquired first); finally, these information is integrated into a structured snapshot timing control parameter set, including target timestamp, confidence, adjustment threshold, and priority information of three snapshot points.

[0081] In step 505, the snapshot timing control parameters are protocol-converted with the camera control interface to generate a final snapshot control instruction sequence containing vehicle head image trigger instructions, vehicle body middle image trigger instructions, and vehicle tail image trigger instructions, wherein the instruction sequence is constructed in the order of snapshot timestamp (vehicle head, vehicle body middle, vehicle tail), and each instruction contains core fields: instruction identifier, trigger timestamp, imaging parameter set, storage configuration, and execution priority; the constructed snapshot control instruction sequence (containing vehicle head, vehicle body middle, and vehicle tail trigger instructions) is sent to the camera control interface, and the instruction details (instruction content, sending time, target timestamp, imaging parameters) are recorded to the system log.

[0082] In a preferred embodiment of the present invention, step 6, controlling the synchronous camera array deployed at the intersection according to the final capture control command sequence, triggers and captures key perspective images of the front, middle, and rear of the vehicle at predetermined time points, including: The camera receives and parses the final capture control command sequence generated in step 5. The parsing process extracts the precise trigger timestamps specified for the capture points at the front, middle, and rear of the vehicle, as well as the imaging parameter set attached to each capture command. These parameters include the focal length, aperture, shutter speed, and ISO sensitivity settings optimized for different capture angles and distances.

[0083] Next, based on the obtained imaging parameters, a pre-configured command is sent to the corresponding camera specified by the command before the predicted capture time. These commands control the camera drive mechanism to adjust the optical lens for precise focusing, locking the focus on the location where the target vehicle is about to appear. At the same time, the exposure time and gain of the camera sensor are dynamically adjusted according to the ambient lighting conditions and preset parameters to ensure that an image with proper exposure and clear details can be obtained at the moment of triggering, thereby completing the active optical and electronic preparation before capture.

[0084] Subsequently, when the precise time point specified by any trigger command in the final capture control command sequence is reached, a trigger signal is immediately sent to the camera associated with that command. This trigger signal instructs the camera to immediately execute high-speed continuous frame image capture in a single frame or a predefined mode to freeze the instant the vehicle reaches the key positions of the front, middle, or rear of the vehicle. Finally, after successfully executing the three trigger commands for the front, middle, and rear of the vehicle in sequence, the synchronous camera array generates a set of vehicle state images captured from different key perspectives, with strict temporal and spatial correspondence. These images are bound to the vehicle's unique identifier, timestamp, and spatial coordinates, forming a high-quality vehicle state image evidence set for traffic incident determination and evidence collection, thus fully realizing closed-loop control from three-dimensional motion prediction to two-dimensional key evidence image capture.

[0085] like Figure 2 As shown, embodiments of the present invention also provide a device for extending the capture distance of a long-range wide-angle electronic police system, comprising: The acquisition module is used to acquire a continuous image sequence of vehicles at traffic light intersections. The continuous image sequence includes images of the front of the vehicle, the middle of the vehicle, and the rear of the vehicle. The continuous image sequence is preprocessed to obtain preprocessed image data. The calculation module is used to calculate the key contour chord length parameters of the vehicle in the imaging plane from the preprocessed image data; construct a perspective projection model based on the key contour chord length parameters and camera calibration parameters; and determine the spatial position data of the reconstructed vehicle based on the perspective projection model. The construction module is configured to construct a pyramid geometry model of the vehicle occupancy space based on the three-dimensional reconstructed vehicle spatial position data, calculate a spatial volume distribution of the vehicle in the monitoring area according to the pyramid geometry model, and generate a three-dimensional space state parameter set containing a vehicle spatial pose, a motion vector and a volume feature; The determination module is configured to calculate vertex coordinates of a vehicle motion path according to the three-dimensional space state parameter set, determine a final imaging timing of the vehicle at a key monitoring point, and generate a final snapshot control instruction sequence containing a vehicle head, a vehicle middle and a vehicle tail. The determination module is configured to calculate vertex coordinates of a vehicle motion path according to the three-dimensional space state parameter set, determine a final imaging timing of the vehicle at a key monitoring point, and generate a final snapshot control instruction sequence containing a vehicle head, a vehicle middle and a vehicle tail.

[0086] As shown in Figure 3 The fixed base flange plate 1 is provided with a waist-shaped long strip mounting hole 2, an extension rod 3 (made of Q235 material square rod), an adjusting mounting clamp 4, an installation bolt 5 and an adjusting reinforcing bar 6. A long-distance wide-angle electronic police extended snapshot viewing distance upgrading method is adopted, a customized support installation scheme is matched, and the law enforcement precision and efficiency are improved.

[0087] The 120-meter long-distance monitoring range of the intersection is covered, and full-profile snapshots of large trucks, cars and other vehicles are realized. The influence of perspective distortion and shielding is eliminated, the traffic violation judgment error rate is reduced to below 3%, the key images of the vehicle head, the vehicle middle and the vehicle tail are accurately captured, and a complete evidence chain is formed. The support system adapts to the 10-level gust environment at the intersection, and ensures the stable operation of the equipment for 24 hours.

[0088] The three-dimensional reconstruction vehicle spatial position error is less than or equal to 0.3 meters; the snapshot timing prediction accuracy is less than or equal to 50 milliseconds, and the key image definition is greater than or equal to 1080P; the support safety factor is greater than or equal to 1.5, and the anti-overturning safety factor is greater than or equal to 2.0; the system continuous operation failure rate is less than or equal to 0.5 times per month.

[0089] After the hardware deployment and support installation are completed, the core parameters are as follows: one wide-angle electronic police snapshot camera, 8mm lens (field of view angle 30°), 9 million pixels, frame rate 25fps; Q235 steel hot-dip galvanized square tube extension rod, length 2.5 meters, fixed with the original electronic police rod horizontal arm through the serrated clamp; one long-focus face cap snapshot camera; 25mm lens, 9 million pixels, responsible for capturing vehicle detail features, independent fixed base, flange plate with waist-shaped mounting hole, supporting 10°-15° angle adjustment, light supplement lamp (2), LED cold light source, light decay ≤5% / year; horizontal distance from camera 1.8 meters, avoiding light interference.

[0090] The embodiment is based on a long-distance wide-angle electronic police system 2.0 version technical solution developed by a traffic police detachment, combined with the existing electric police pole upgrading scene, and specifically implements the method of extending the shooting distance of the long-distance wide-angle electronic police, and the following is the detailed implementation process: Step 1: Obtain a continuous image sequence of a vehicle at a red light intersection The embodiment is aimed at the intersection scene where the electric police pole is more than 23 meters away from the stop line, or the pole is too close to the stop line and cannot effectively collect the illegal intersection of large and small vehicles. The existing electric police pole, extended support and short focal length lens device deployment scheme is adopted: 1. Device configuration: install a customized shooting extension bracket (Q235 steel material hot-dip galvanized square tube, length 2-4.67 meters calculated by formula according to shooting requirements) on the existing electric police pole cross arm, replace the original camera lens with an 8mm short focal length lens (field of view about 30°), retain the original bayonet and light supplement equipment, and form a synchronous camera array. Among them, the 8mm short focal length lens is responsible for expanding the shooting field of view, the original light supplement equipment and the new bracket maintain a horizontal distance of ≥1.5 meters to avoid light interference.

[0091] 2. On-site adaptation: Before installation, the professional personnel assess the load-bearing capacity of the electric police pole to ensure that the support load-bearing capacity is ≥1.2 times the total weight of the equipment (including camera, protective cover, line, etc.), and there is no obstruction such as branches, billboards in the extension direction. The vertical distance between the cross arm and the ground is not less than 6 meters, and the road clearance meets the traffic requirements.

[0092] 3. Image acquisition: The camera array synchronously acquires video streams at a preset high frame rate, identifies moving targets in real time through vehicle detection algorithms, and judges the vehicle type (distinguishes between large trucks 15.5m, extended axle trucks 19m, etc.). When the vehicle enters the preset trigger area, the multi-frame continuous acquisition mode is started, and the continuous image sequence of the vehicle head, the middle of the vehicle body and the tail is captured in time sequence, the timestamp, camera spatial coordinates (including three-dimensional position after the extension of the extended support) and vehicle unique identifier of each image are recorded synchronously, and stored in the buffer area.

[0093] Step 2: Preprocess the continuous image sequence 1. Image screening and correction: First, eliminate low-quality image frames with motion blur, overexposure / underexposure, and retain valid images that meet the clarity threshold. For the optical properties of the 8mm lens, a geometric correction algorithm is used to correct the lens edge distortion, ensuring that the target falls within the 1 / 2 range of the center of the picture, improving the subsequent recognition accuracy.

[0094] 2. Standardization: Convert the corrected image to a standard color space, adjust the brightness and contrast to a pre-set range, and eliminate the influence of light differences at different times. Use background difference method and image segmentation technology to extract the vehicle foreground area, remove the road background, surrounding buildings and other interference information, and generate a binary mask image.

[0095] 3. Image cropping and reorganization: After registering the mask image with the original image, crop the image data containing only the vehicle target and standardize the size, and reorganize it into a standardized image sequence in chronological order to prepare for subsequent feature extraction.

[0096] Step 3: Three-dimensional reconstruction of vehicle spatial position data (1) Calculate the vehicle key contour chord length parameter 1. Edge detection and closed contour extraction: Perform gradient amplitude edge detection operator on the pre-processed standardized image, extract the closed boundary curve of the vehicle contour through non-maximum suppression and double-threshold hysteresis connection algorithm, and ensure that the contour is uninterrupted.

[0097] 2. Ellipse model fitting: Uniformly sample 200 contour points from the closed boundary curve, calculate the geometric center as the ellipse center point, and determine the ellipse major axis direction (consistent with the vehicle longitudinal direction) and major and minor axis parameters through eigenvalue decomposition of the covariance matrix. The major axis corresponds to the vehicle longitudinal length direction, and the minor axis corresponds to the transverse width direction.

[0098] 3. Local chord length sampling: Divide the vehicle contour into 14 equal length intervals along the major axis direction, set 15 sampling sections, and calculate the Euclidean distance between the intersection points as the local chord length. Record the major axis position coordinates, intersection point coordinates and local chord length values of each sampling point to form the local chord length sampling data set.

[0099] 4. Global chord length feature vector generation: Normalize the local chord length based on the ellipse minor axis length, and normalize the sampling point position based on the major axis length. Extract the width distribution statistical features (maximum, minimum, mean, standard deviation), contour curvature variation features (first-order / difference sequence), and regional morphological features (head, body, tail partition parameters), and splice them to form the global chord length feature vector.

[0100] (2) Construct a perspective projection model 1. Camera intrinsic parameter acquisition: Obtain the camera intrinsic parameter matrix through Zhang Zhengyou's chessboard calibration method, which includes the horizontal / vertical focal length of an 8mm lens, the principal point coordinates, and the radial distortion coefficients (k1, k2, k3) and tangential distortion coefficients (p1, p2). Calibrate the influence of lens optical distortion.

[0101] 2. Coordinate mapping relationship establishment: Perform tensor product operation on the global chord length feature vector and the camera intrinsic matrix to generate a high-dimensional coupled tensor, and obtain a 10-dimensional compressed feature vector through principal component analysis dimension reduction. Combine the extension support installation parameters (original height of electric pole H, horizontal length of extension pole AL), and calculate the composite scale factor of each image point through spatial gridding processing and bilinear interpolation to establish the mapping relationship from the image plane to the camera coordinate system.

[0102] 3. Perspective transformation matrix optimization: Select 12 ground calibration control points at the intersection (measure the world coordinates through high-precision differential GPS), convert the image coordinates to the camera coordinate system after distortion correction, and solve the initial perspective transformation matrix using the direct linear transformation algorithm. Suppress noise through singular value decomposition and adjust the coplanar constraint of the 5 grounding points at the bottom of the vehicle to finally generate a perspective projection model that integrates the support installation parameters.

[0103] (3) Determine the three-dimensional spatial position data 1. Prior model registration: Load a three-dimensional geometric model library containing large trucks, small cars and other vehicle types, match the closest vehicle model based on the global chord length feature vector, project the model to the image plane, evaluate the alignment quality through Hausdorff distance, and adjust the initial pose of the model to make the center coincide with the image center of the vehicle.

[0104] 2. Pose parameter optimization: Use the iterative closest point algorithm to match and optimize the model point set and the scene point set obtained by image back projection, introduce vehicle kinematic constraints (translation ≤ 0.2 meters / iteration, rotation angle ≤ 5° / iteration) and ground height constraints (Z coordinate 0.1-0.3 meters), and solve to obtain a 3x3 rotation matrix and a 3x1 translation vector.

[0105] 3. Three-dimensional point cloud generation and processing: Extract 50 key feature points of the vehicle (front end of the vehicle, rear end of the vehicle, grounding points of the wheels, etc.), convert them to the world coordinate system through the pose parameters, fuse the image back projection coordinates to obtain a three-dimensional point cloud, and generate a sparse point cloud data set of 2000 points through outlier detection and interpolation processing.

[0106] 4. Construction of spatial position data set: Convert the point cloud into a spatial occupancy grid with a resolution of 0.1 meters, calculate the vehicle center point coordinates, spatial orientation angles (yaw angle, pitch angle, roll angle), and circumscribed cube dimensions (large truck length 15.5m / 19m, width 2.5m, height ≥ 3.2m), and integrate them into a three-dimensional spatial position data set containing timestamps and confidence indicators.

[0107] Step 4: Construct a pyramid geometric model and generate a three-dimensional spatial state parameter set (1) Construct a pyramid geometric model of the vehicle occupancy space 1. Pyramid apex determination: Extract the vehicle motion direction vector (calculated by smoothing the coordinates of the center points of three consecutive frames) and the current speed based on the three-dimensional spatial position data, dynamically adjust the preset safety distance (10 meters for large trucks, increase by 2 meters for every 10 km / h increase in speed), extend the distance along the motion direction to obtain the pyramid apex, and determine the final coordinates after correction of the road curvature and slope.

[0108] 2. Pyramid base generation: Extract 80 key feature points from the bottom of the vehicle, project them onto the ground plane (Z = 0.15 meters, considering the ground clearance), perform convex hull processing through DBSCAN clustering and Graham scan algorithm, and generate a regular polygon with 12-16 vertices as the pyramid base.

[0109] 3. Closed pyramid model construction: Connect the pyramid apex with each vertex of the base polygon to form a closed pyramid geometric model, generate side facets through triangulation, apply height constraints (≤ 3.5 meters for large trucks) and contour fitting optimization to ensure that the model includes the vehicle body and the front safety buffer area.

[0110] (2) Calculate the spatial volume distribution and generate a set of three-dimensional spatial state parameters 1. Horizontal slice sectioning: Divide the pyramid into 25 layers (adapted for large vehicles) along the height direction, with each layer having a thickness of ≤ 0.2 meters, and calculate the intersection line of each layer with the pyramid side to obtain the cross-sectional contour polygon.

[0111] 2. Grid occupancy probability calculation: Grid each slice layer with a resolution of 0.05-0.15 meters, calculate the coincidence degree of the grid and the contour using the ray intersection method, and combine the point cloud density and motion state weight to obtain the spatial occupancy probability density P(i,j,k).

[0112] 3. Volume distribution feature extraction: Calculate the volume barycentric coordinates, three-dimensional direction variance, spatial occupancy rate, and effective volume of the pyramid model, where the spatial occupancy rate of large trucks must be ≥ 0.7 to ensure no severe occlusion.

[0113] 4. Spatio-temporal fusion: Fuse the volume distribution features with the three-dimensional spatial position data set to integrate the vehicle real-time spatial attitude angle (corrected by the main direction of the volume distribution), three-dimensional motion vector (with volume change rate as an acceleration correction factor), volume features, and spatial occupancy state to generate a structured three-dimensional spatial state parameter set.

[0114] Step 5: Determine the snapshot opportunity and generate control instruction sequence (1) Calculate the vehicle motion path apex coordinates 1. Trajectory Modeling: Extract 15 frames of historical trajectory point sequences from the 3D spatial state parameter set (after outlier removal and interpolation compensation). Using the vertical position Y as the independent variable and the horizontal position X as the dependent variable, fit the quadratic parabola equation using the weighted least squares method. The weighting coefficient for recent trajectory points is . (n is the frame number).

[0115] 2. Vertex calculation: using formulas Calculate the vertical coordinates of the vertex of the parabola, and substitute them into the equation to obtain the horizontal coordinates. Determine vertical coordinates by combining road slope information ( (where α is the current elevation and α is the road slope angle). This vertex serves as a key turning point in the vehicle's trajectory, and after being corrected by the lane centerline projection, it ensures conformity to the road direction.

[0116] (2) Determine the coordinates of key capture points 1. Capture point offset setting: Based on the stop line position in the high-precision map of the intersection ( The dynamic adjustment of the capture point offset distance is as follows: the capture point at the front of the truck is 6.0 meters away from the stop line, the capture point in the middle of the vehicle is located L / 2 behind the capture point at the front of the truck (L=15.5m or 19m), and the capture point at the rear of the truck is located L behind the capture point at the front of the truck.

[0117] 2. Trajectory-aligned positioning: The longitudinal coordinates of the vehicle front capture point are... Substituting into the equation of the parabola, we obtain the horizontal coordinate. Combined with road slope calculation Similarly, calculate the capture point in the middle of the vehicle body ( ) and rear-end camera points ( Curvature compensation ensures that the capture point is located on the vehicle's center trajectory and within the camera's effective field of view (horizontal ±30°, vertical ±15°).

[0118] (3) Accurate calculation of capture distance and extension rod length (supplementary formula derivation and verification) Core parameter definition, Indicates the lens focal length (mm, which remains constant); Indicates the original height (m) of the monitoring equipment on the electric warning pole from the ground. This indicates the horizontal length of the extension pole (m, which is the length of the horizontal arm of the electric warning pole extending outwards). This indicates the capture distance (m) without the extension pole, i.e., the original electronic police pole to the vehicle capture distance. This indicates the capture distance (m, new capture distance) after adding the extension rod. This indicates the length of the vehicle being photographed (m, 15.5m for large trucks and 19m for extended axle trucks). represents the comprehensive coefficient (determined by the lens focal length, device height, etc., which needs to be calibrated on site); represents the standard reference focal length (mm, same type as the actual lens, such as 16mm or 8mm).

[0119] 2. Derivation and verification of snapshot distance calculation formula Law extraction: under the same length of extension rod, the snapshot distance increment is proportional to the original snapshot distance , and is related to the lens focal length. Calculate the proportional coefficient through field measurement data: 16mm lens: , , , , the proportional coefficient ; 8mm lens: , , , , the proportional coefficient .

[0120] General formula derivation: combined with the principle of similar triangles, is inversely proportional to the focal length ( , is a constant), and the comprehensive coefficient is introduced to calibrate the deviation. The final snapshot distance formula is: ; Case verification: 16mm lens ( , , , , ): (consistent with the measured value); 8mm lens ( , , , , ): (consistent with the measured value).

[0121] 3. Extension rod length calculation formula (satisfy the target snapshot distance requirement) Formula transformation: based on the snapshot distance formula, when , the extension rod length calculation formula is simplified as: ; Case application: 16mm lens target snapshot distance (satisfy the length of large truck), , : 8mm lens target snapshot distance , , : .

[0122] (4) Calculate the snapshot time prediction value According to the current longitudinal speed (smoothed by Kalman filtering) and acceleration of the vehicle, the time to reach each snapshot point is calculated: in uniform speed state (absolute value of acceleration <0.2 m / s²), the effective distance is directly divided by the speed; in acceleration / deceleration state, the time is calculated by combining the acceleration; in turning state (change rate of yaw angle >3° / frame), the speed is reduced by 10%-20% before calculating the time. The time prediction value of each snapshot point is smoothed by weighted average of historical motion trend to generate millisecond-level timestamp.

[0123] (5) Generate control instruction sequence Convert the snapshot timing control parameters to the camera control interface to generate the final snapshot control instruction sequence containing three trigger instructions, arranged in chronological order: Vehicle head image trigger instruction: instruction identification, trigger timestamp, imaging parameters (8mm lens focal length, optimized aperture / shutter speed), storage configuration, execution priority (highest); vehicle middle image trigger instruction: core fields are the same as before, execution priority is second; vehicle tail image trigger instruction: core fields are the same as before, execution priority is third. The instruction sequence is sent to the camera control interface at the same time, and the instruction details are recorded to the system log.

[0124] Step 6: Control the synchronous camera array to capture key view images 1. Pre-configuration preparation: After the camera receives the control instruction sequence, it completes the pre-configuration before the trigger timestamp: adjusts the lens focus to the preset position of the snapshot point according to the imaging parameters, dynamically adjusts the sensor exposure time and gain based on the ambient light, and the lighting equipment enters the standby trigger state.

[0125] 2. Precise trigger snapshot: When the system time reaches the trigger timestamp set by each instruction, the camera array is triggered in turn to capture key view images of the vehicle head, middle, and tail, freezing the instantaneous state of the vehicle at the corresponding position.

[0126] 3. Image storage and binding: Store the captured images with the vehicle unique identifier, timestamp, and spatial coordinate information. The images must meet the requirements of clear identification of truck license plate, vehicle model, and driving state, without blind spots or distortion, and meet the complete standard of non-field law enforcement snapshot elements.

[0127] 1. Snapshot distance calculation formula (same lens focal length) ; Parameter description: The capture distance (m) is the original pole without extension. The horizontal length of the extension rod (m); This is the actual lens focal length (mm). This is the standard reference focal length (mm, the same type as the actual lens). The composite coefficient (for a 16mm lens) 8mm lens ).

[0128] 2. Formula for calculating the length of the extension rod (assuming the same lens focal length and the target capture distance is met) ; Parameter description: The target capture distance (m, which must cover the length of the large truck). The original pole capture distance (m); This is a composite coefficient (calibrated on-site based on lens focal length).

[0129] 3. Calculation formula for the limitation of extended support length (wind resistance and bending stress constraints) (1) Definition of basic formula Wind force calculation formula (corrected aerodynamic coefficient): ; Parameter description: Wind force (N); air density (taken under standard conditions) ); Wind speed (m / s); For the projected area ( ); The drag coefficient (taken for octagonal rod) ).

[0130] Bending moment calculation formula: ; Parameter description: The bending moment is (N·m). The height of the point of action of the wind force (m). ( (Height of the traffic enforcement pole).

[0131] Bending stress calculation formula: ; Parameter description: Bending stress (Pa); The section modulus of bending ( According to the outer diameter of the octagonal rod , wall thickness Calculation , is the inner diameter.

[0132] (2) Length limit derivation formula ; Parameter description: is the allowable stress of steel material ); (Eight bar projection area approximate calculation); Actual application needs to be substituted into specific rod specifications, design wind speed, etc. to calculate the limit value.

[0133] This embodiment realizes effective snapping of illegal behavior of large trucks at long-distance intersections of more than 23 meters by installing customized extension supports and optimizing lens configuration, combined with precise geometric modeling and trajectory prediction, without moving the electric police stand. The production cost of a single support is only about 500 yuan, which saves more than 96% of the traditional moving rod modification, and the construction does not require a crane and can be implemented by maintenance personnel, shortening the construction period from 10-15 days to a single day. At the same time, through standardized formula calculation and three-dimensional reconstruction technology, the illegal judgment error rate is greatly reduced, and the modernization level of road traffic management is improved.

[0134] Through multiple tests, a certain information management team determines that the long-distance wide-angle snapping support is suitable for electric police poles that are more than 23 meters away from the stop line, or for city road construction when the electric police pole is too close to the stop line (less than 18 meters) and cannot collect large vehicle violations. The site has space conditions for adding extension supports, such as no obstructions above the pole, no impact on vehicle and pedestrian traffic, the horizontal arm of the national and provincial road point is not less than 6 meters vertically from the ground, etc.

[0135] Before installing the support, the construction personnel first need to conduct on-site investigation in advance to confirm the existing electric police pole bearing capacity, and it is recommended to be evaluated by professional personnel on site to avoid overloading the pole due to the addition of the support. Secondly, the diameter of the electric police pole at the support installation position needs to be determined, and whether there are obstructions in the rear direction (such as branches, billboards, signboards, etc.) needs to be investigated. If necessary, clean up in advance to ensure that it does not affect the road clearance.

[0136] (II) Support design parameters 1. Snapping support specifications (1) Support material: hot-dipped galvanized square tube, corrosion-resistant and high-strength, size set according to actual bearing requirements. (2) Support length: calculated according to the snapping blind area distance, usually 1.5-3 meters, can be measured according to the intersection scene to ensure that the monitoring equipment is directly opposite the large truck track after extension. (3) Connection method: use installation hoops and bolts to fix with the original electric police pole horizontal arm to ensure connection stability, wind resistance level not less than the original electric police pole standard.

[0137] 2. Bracket and equipment fixation (1) Design requirements: custom-made bracket flange plate, welded with the end of the extension rod, and the surface is treated with hot galvanizing anti-rust treatment according to requirements. The flange plate reserves the installation hole position of the monitoring camera, ensuring that the angle can be adjusted during on-site installation of the equipment. (2) Manufacturing process: the welding joint is full welding, and the surface and welding seam should not have the phenomena of incomplete penetration, slag inclusion, and incomplete filling of craters. Manual arc welding is used for short welds or complex parts. After cutting, the steel parts are deburred and sanded to expose a glossy surface to avoid impurities affecting the welding quality. After welding, the appearance is inspected, and the welding seam surface is free of defects such as pores, slag inclusion, cracks, and incomplete penetration. (3) Overall appearance: if necessary, the extension rod and bracket can be spray painted according to the original traffic pole color to maintain uniformity with the appearance of road facilities.

[0138] (Three) Bracket testing process 1. Preliminary design A grab shot extension bracket is added to the middle of the 13-meter traffic pole at the test intersection, which is fixed by eight-rib custom-made clamps and bolts. The perpendicular distance between the extension rod and the traffic pole is 1 meter, the height is raised by 0.5 meters, the length of the extension rod is about 1.18 meters, and the design is as follows. Figure 3

[0139] First on-site test feedback: According to the preliminary test, after the height of the extension rod is raised, the grab shot range is blocked by the pole, and the vehicle violation cannot be effectively captured. Moreover, the clamp size at the joint between the extension rod and the traffic pole is fixed and cannot be adjusted according to the diameter of the pole, which needs to be further improved.

[0140] 2. Second adjustment design After on-site testing and surveying, the perpendicular distance between the extension rod and the traffic pole is adjusted to about 2 meters, with a parallel drop of 0.4 meters. The joint between the extension rod and the traffic pole is adjusted to a crocodile tooth installation, which can be adjusted according to the diameter of the pole.

[0141] Second on-site test feedback: The grab shot range test effect is good, with the grab shot angle distance improved from 15 meters to 19 meters. However, the following problems are found: (1) The horizontal arm where the camera is installed is welded on one side, and the crocodile tooth clamp on the other side is too small, causing it to not fit tightly with the traffic pole. It is recommended to use a clamp for the eight-rib rod and a crocodile tooth for the round rod. (2) The bracket is too heavy after adding the camera, which has a risk of falling. To avoid collisions with over-height vehicles, it is recommended to consider posting reflective strips or installing warning lights at the back, which can take power from the camera for warning and prompting of over-limit vehicles, which needs to be further improved.

[0142] 3. Third edition adjustment design ​In view of the first two field test, the extension rod and the electric police eight prong rod are still connected by the hoop connection. In order to ensure that the extension support is installed and clamped with the electric police rod, the welding fastening piece is increased at the installation hoop position and the screw hole is increased.

[0143] Third field test feedback: The extension rod and the electric police rod installation hoop welding fastening piece fixing position does not match the actual rod diameter, resulting in the installation of the extension support not being very firm, the hardness of the hoop is not enough, the hoop is easy to deform after tightening the screw, and further improvement is needed according to the field feedback.

[0144] 4. Fourth edition adjustment design In view of the feedback of the field test results, the extension support installation hoop is adjusted in detail. According to the installation position and specific size of the rod, the hoop piece is adjusted and the welding reinforcement is increased.

[0145] Fourth field test feedback: The installation effect is good, and the electric police short distance lens is replaced, the grab shot visual angle distance is extended from the original 11 meters to 21 meters. In the case of the distance between the intersection electric police rod and the stop line being 23.5 meters, the test data is: (1) The electric police uses a 16mm lens, the grab shot visual distance is 11 meters behind the stop line; (2) The electric police uses a 16mm lens, after adding the extension support, the grab shot visual distance is 15 meters behind the stop line; (3) The electric police replaces an 8mm lens, the grab shot visual distance is 18 meters behind the stop line; (4) The electric police replaces an 8mm lens, after adding the extension support, the grab shot visual distance is 21 meters behind the stop line.

[0146] The test intersection grab shot visual distance comparison diagram is shown in Figure 4 Through actual measurement, it is proved that the addition of the electric police grab shot extension support can effectively solve the problem of large truck illegal camera grab shot blind area, and effectively improve the non-site law enforcement efficiency of the intersection traffic monitoring equipment. The electric police grab shot effect comparison picture is as follows Figure 5 , the electric police truck picture before installing the extension support and replacing the lens. The electric police truck picture after installing the extension support and replacing the 8mm lens is shown in Figure 6 ; the electric police truck picture after installing the extension support and replacing the 8mm lens is shown in Figure 7 ; the actual illegal evidence picture of the electric police truck after installing the extension support is shown in Figure 8 ; In summary, the long-distance wide-angle electric police snapshot extension bracket developed by Zibo Traffic Management Brigade truly solves the technical problem that has long plagued the traffic police department, i.e., the electric police snapshot distance is not enough to effectively snapshot the illegal truck, expands the snapshot efficiency of the electronic police in the field of traffic management, and has higher practical popularization value. Through large-scale testing of the traffic police in the whole city, multiple specifications of the bracket suitable for different intersection scenes, different rod sizes and different snapshot distances are developed and popularized, the snapshot quality and efficiency of the non-site law enforcement equipment in the whole city are improved, the occurrence of major traffic accidents in the jurisdiction is prevented and reduced, and the road traffic safety control level is improved.

[0147] It should be noted that the device is a system corresponding to the above method, and all implementation manners in the above method embodiments are applicable to this embodiment and can also achieve the same technical effects.

Claims

1. A method for extending the capture distance of a long-range wide-angle electronic police system, characterized in that, The method includes: Acquire a continuous image sequence of vehicles at a traffic light intersection. The continuous image sequence includes images of the front of the vehicle, the middle of the vehicle, and the rear of the vehicle. A continuous image sequence is preprocessed to obtain preprocessed image data; Calculate the key contour chord length parameters of the vehicle in the imaging plane from the preprocessed image data; construct a perspective projection model based on the key contour chord length parameters and camera calibration parameters; determine the spatial position data of the reconstructed vehicle based on the perspective projection model. Based on the 3D reconstructed vehicle spatial location data, a pyramidal geometric model of the space occupied by the vehicle is constructed; the spatial volume distribution of the vehicle in the monitoring area is calculated according to the pyramidal geometric model, and a 3D spatial state parameter set containing the vehicle's spatial attitude, motion vector and volume features is generated. Based on the three-dimensional spatial state parameter set, calculate the vertex coordinates of the vehicle's motion path, determine the final imaging timing of the vehicle at key monitoring points, and generate a final capture control command sequence that includes the front, middle, and rear of the vehicle. Based on the final capture control command sequence, the synchronous camera array deployed at the intersection is controlled to trigger and capture key perspective images of the front, middle, and rear of the vehicle at predetermined time points.

2. The method for extending the capture distance of a long-range wide-angle electronic police system according to claim 1, characterized in that, Calculate the key contour chord length parameters of the vehicle in the imaging plane from the preprocessed image data, including: An edge detection operator is performed on the preprocessed image data to extract the closed boundary curve of the vehicle contour; The closed boundary curve is fitted to an elliptical geometric model, and the principal axis direction and major and minor axis parameters of the ellipse are determined. Multiple equally spaced sampling points are selected on the vehicle profile along the main axis direction, and the Euclidean distance between adjacent sampling points is calculated as the local chord length. The local chord length is associated with the elliptical geometric parameters to generate a global chord length feature vector that characterizes the vehicle's contour shape. The global chord length feature vector includes the vehicle's width distribution, length ratio, and contour curvature variation characteristics on the imaging plane.

3. The method for extending the capture distance of a long-range wide-angle electronic police system according to claim 2, characterized in that, Based on the key contour chord length parameters, a perspective projection model is constructed in conjunction with camera calibration parameters, including: Obtain a pre-calibrated camera intrinsic parameter matrix, which includes focal length parameters, principal point coordinates, and lens distortion coefficients; The global chord length eigenvector is multiplied by the camera intrinsic parameter matrix using a tensor product to establish a mapping relationship between the image plane and the camera coordinate system. Based on the mapping relationship from the image plane to the camera coordinate system, and combined with the ground calibration control points pre-set at the traffic light intersection, the initial perspective transformation matrix between the image plane and the world coordinate system is calculated. Singular value decomposition is performed on the initial perspective transformation matrix to obtain three orthogonal matrices and singular value vectors; the distribution characteristics of the singular value vectors are suppressed to obtain noise-suppressed orthogonal matrices; based on the noise-suppressed orthogonal matrices, the optimized perspective transformation matrix is ​​reconstructed. The reconstructed and optimized perspective transformation matrix is ​​fused with the geometric constraints of the vehicle contour to obtain the adjusted perspective transformation matrix. The adjusted perspective transformation matrix is ​​concatenated with the mapping relationship from the image plane to the camera coordinate system to generate the final perspective projection model.

4. A method for extending the capture distance of a long-range wide-angle electronic police system according to claim 3, characterized in that, Based on the perspective projection model, determine the spatial location data of the reconstructed 3D vehicle, including: The perspective projection model is registered and aligned with the preset vehicle prior geometry model, which includes the length, width and height ratio constraints and contour topology of a standard vehicle. The matching error between the vehicle's prior geometric model and perspective projection model is optimized by iterative nearest point algorithm, and the vehicle's pose parameters in the world coordinate system are calculated. The pose parameters include translation vectors and rotation matrices. Based on the pose parameters, the key feature points on the vehicle outline are back-projected in three dimensions to generate three-dimensional point cloud data of the vehicle within the monitoring area. The three-dimensional point cloud data is converted into a spatial occupancy grid, and the coordinates of the vehicle's center point, spatial orientation angle, and circumscribed cube size are calculated to form a three-dimensional spatial position dataset containing the vehicle's precise spatial position, direction of movement, and physical dimensions.

5. A method for extending the capture distance of a long-range wide-angle electronic police system according to claim 4, characterized in that, Step 4: Based on the 3D reconstructed vehicle spatial location data, construct a pyramidal geometric model of the space occupied by the vehicle, including: Based on the vehicle's motion direction vector and current speed parameters, the spatial position of the pyramid vertex is determined along the extension line of the motion direction. The pyramid vertex is located at a preset safe distance in front of the vehicle, and the preset safe distance is dynamically adjusted according to the vehicle type and speed. Key feature points on the vehicle outline are projected onto the ground plane of the traffic light intersection to form a polygonal bottom outline of the vehicle; the convex hull of the polygonal bottom outline of the vehicle is processed to generate a regular polygon with a pyramidal base. A closed pyramidal geometric model is constructed by connecting the vertices of the pyramid to the vertices of the regular polygons on the base of the pyramid. The pyramidal geometric model represents the three-dimensional spatial range occupied by the vehicle within the monitoring area, including the spatial volume of the vehicle body and the front safety buffer zone.

6. A method for extending the capture distance of a long-range wide-angle electronic police system according to claim 5, characterized in that, Based on the pyramidal geometric model, the spatial volume distribution of the vehicle within the monitored area is calculated, generating a three-dimensional spatial state parameter set containing the vehicle's spatial attitude, motion vectors, and volumetric features, including: The geometric model of the pyramid is spatially partitioned, and the pyramid is divided into multiple horizontal slice layers along the height direction. Each horizontal slice layer is meshed, and the spatial occupancy probability density of each mesh cell is calculated. The spatial occupancy probability density is weighted according to the degree of overlap between the mesh cell and the actual contour of the vehicle. Based on the spatial occupancy probability density of each grid cell, the overall volume distribution characteristics of the pyramidal geometric model are calculated, including the volume centroid coordinates, volume distribution variance, and spatial occupancy rate. By spatiotemporally fusing volume distribution features with a 3D spatial location dataset, a 3D spatial state parameter set is generated, which includes the vehicle's real-time spatial attitude angles, 3D motion vectors, volume distribution features, and spatial occupancy status.

7. A method for extending the capture distance of a long-range wide-angle electronic police system according to claim 6, characterized in that, Based on the three-dimensional spatial state parameter set, the vertex coordinates of the vehicle's motion path are calculated to determine the final imaging timing of the vehicle at key monitoring points. A final capture control command sequence is generated, including the front, middle, and rear of the vehicle, comprising: Extract the vehicle's three-dimensional motion vectors, spatial attitude angles, and historical trajectory point sequences from the three-dimensional spatial state parameter set; Based on the historical trajectory point sequence and three-dimensional motion vector, the parabolic equation parameters of the vehicle's motion trajectory are calculated. The parabolic equation parameters include quadratic coefficients, linear coefficients, and constant terms. Based on the parabola equation parameters, calculate the vertex coordinates of the parabola trajectory. The vertex coordinates correspond to the highest or lowest point of the vehicle's trajectory, representing the key turning points of the vehicle's motion. Based on the stop line position and vehicle type parameters at the traffic light intersection, determine the spatial coordinates of the front, middle, and rear camera points on the parabolic trajectory. Based on the vehicle's current speed, acceleration, and spatial attitude angle, calculate the predicted time of the vehicle's arrival at each capture point, and generate capture timing control parameters including timestamps. The capture timing control parameters are converted into a protocol with the camera control interface to generate the final capture control command sequence, which includes trigger commands for the front of the vehicle, the middle of the vehicle, and the rear of the vehicle.

8. A method for extending the capture distance of a long-range wide-angle electronic police system according to claim 7, characterized in that, Based on the final capture control command sequence, the synchronous camera array deployed at the intersection is controlled to trigger and capture key perspective images of the front, middle, and rear of the vehicle at predetermined time points, including: Receive and parse the final capture control command sequence to obtain the parsed imaging parameters; Based on the obtained imaging parameters, each camera is controlled to adjust its focus, exposure and sensitivity in advance. When the time reaches the time point specified by any trigger command, a trigger signal is sent to the corresponding camera to control the camera to perform single-frame or high-speed continuous frame image capture and generate a vehicle status image evidence set.

9. A method for extending the capture distance of a long-range wide-angle electronic police system according to claim 8, characterized in that, The three-dimensional spatial state parameter set includes the vehicle's real-time spatial attitude angles, three-dimensional motion vectors, volume distribution characteristics, and spatial occupancy status.

10. A device for extending the capture distance of a long-range wide-angle electronic police system, the device implementing the method as described in any one of claims 1 to 9, characterized in that, include: The acquisition module is used to acquire a continuous image sequence of a vehicle at a traffic light intersection. The continuous image sequence includes images of the front of the vehicle, the middle of the vehicle, and the rear of the vehicle. A continuous image sequence is preprocessed to obtain preprocessed image data; The calculation module is used to calculate the key contour chord length parameters of the vehicle in the imaging plane from the preprocessed image data; construct a perspective projection model based on the key contour chord length parameters and camera calibration parameters; and determine the spatial position data of the reconstructed vehicle based on the perspective projection model. The module is used to construct a pyramidal geometric model of the space occupied by the vehicle based on the 3D reconstructed vehicle spatial location data; calculate the spatial volume distribution of the vehicle in the monitoring area based on the pyramidal geometric model, and generate a 3D spatial state parameter set containing the vehicle's spatial attitude, motion vector and volume features. The determination module is used to calculate the vertex coordinates of the vehicle's motion path based on the three-dimensional spatial state parameter set, determine the final imaging timing of the vehicle at key monitoring points, and generate a final capture control command sequence that includes the front, middle and rear of the vehicle. The determination module is used to control the synchronous camera array deployed at the intersection according to the final capture control command sequence, triggering and capturing key perspective images of the front, middle and rear of the vehicle at predetermined time points.

Images